add extra meaning to WebPDecBuffer::is_external_memory

If value is '2', it means the buffer is a 'slow' one, like GPU-mapped memory.

This change is backward compatible (setting is_external_memory to 2
will be a no-op in previous libraries)

dwebp:  add flags to force a particular colorspace format
new flags is:
   -pixel_format {RGB,RGBA,BGR,BGRA,ARGB,RGBA_4444,RGB_565,
                  rgbA,bgrA,Argb,rgbA_4444,YUV,YUVA}
and also,external_memory {0,1,2}
These flags are mostly for debuggging purpose, and hence are not documented.

Change-Id: Iac88ce1e10b35163dd7af57f9660f062f5d8ed5e

examples/dwebp.c

@@ -67,8 +67,13 @@ typedef enum {
   PGM,
   BMP,
   TIFF,
-  YUV,
+  RAW_YUV,
-  ALPHA_PLANE_ONLY  // this is for experimenting only
+  ALPHA_PLANE_ONLY,  // this is for experimenting only
+  // forced colorspace output (for testing, mostly)
+  RGB, RGBA, BGR, BGRA, ARGB,
+  RGBA_4444, RGB_565,
+  rgbA, bgrA, Argb, rgbA_4444,
+  YUV, YUVA
 } OutputFileFormat;
 
 #ifdef HAVE_WINCODEC_H
@@ -175,7 +180,7 @@ static int WritePNG(const char* out_file_name, int use_stdout,
   const uint32_t height = buffer->height;
   uint8_t* const rgb = buffer->u.RGBA.rgba;
   const int stride = buffer->u.RGBA.stride;
-  const int has_alpha = (buffer->colorspace == MODE_BGRA);
+  const int has_alpha = WebPIsAlphaMode(buffer->colorspace);
 
   return SUCCEEDED(WriteUsingWIC(out_file_name, use_stdout,
                                  MAKE_REFGUID(GUID_ContainerFormatPng),
@@ -193,7 +198,7 @@ static int WritePNG(FILE* out_file, const WebPDecBuffer* const buffer) {
   const uint32_t height = buffer->height;
   uint8_t* const rgb = buffer->u.RGBA.rgba;
   const int stride = buffer->u.RGBA.stride;
-  const int has_alpha = (buffer->colorspace == MODE_RGBA);
+  const int has_alpha = WebPIsAlphaMode(buffer->colorspace);
   volatile png_structp png;
   volatile png_infop info;
   png_uint_32 y;
@@ -259,6 +264,24 @@ static int WritePPM(FILE* fout, const WebPDecBuffer* const buffer, int alpha) {
   return 1;
 }
 
+// Save 16b mode (RGBA4444, RGB565, ...) for debugging purpose.
+static int Write16bAsPGM(FILE* fout, const WebPDecBuffer* const buffer) {
+  const uint32_t width = buffer->width;
+  const uint32_t height = buffer->height;
+  const uint8_t* const rgba = buffer->u.RGBA.rgba;
+  const int stride = buffer->u.RGBA.stride;
+  const uint32_t bytes_per_px = 2;
+  uint32_t y;
+
+  fprintf(fout, "P5\n%u %u\n255\n", width * bytes_per_px, height);
+  for (y = 0; y < height; ++y) {
+    if (fwrite(rgba + y * stride, width, bytes_per_px, fout) != bytes_per_px) {
+      return 0;
+    }
+  }
+  return 1;
+}
+
 static void PutLE16(uint8_t* const dst, uint32_t value) {
   dst[0] = (value >> 0) & 0xff;
   dst[1] = (value >> 8) & 0xff;
@@ -271,7 +294,7 @@ static void PutLE32(uint8_t* const dst, uint32_t value) {
 
 #define BMP_HEADER_SIZE 54
 static int WriteBMP(FILE* fout, const WebPDecBuffer* const buffer) {
-  const int has_alpha = (buffer->colorspace != MODE_BGR);
+  const int has_alpha = WebPIsAlphaMode(buffer->colorspace);
   const uint32_t width = buffer->width;
   const uint32_t height = buffer->height;
   const uint8_t* const rgba = buffer->u.RGBA.rgba;
@@ -332,7 +355,7 @@ static int WriteBMP(FILE* fout, const WebPDecBuffer* const buffer) {
 #define TIFF_HEADER_SIZE (EXTRA_DATA_OFFSET + EXTRA_DATA_SIZE)
 
 static int WriteTIFF(FILE* fout, const WebPDecBuffer* const buffer) {
-  const int has_alpha = (buffer->colorspace != MODE_RGB);
+  const int has_alpha = WebPIsAlphaMode(buffer->colorspace);
   const uint32_t width = buffer->width;
   const uint32_t height = buffer->height;
   const uint8_t* const rgba = buffer->u.RGBA.rgba;
@@ -418,7 +441,7 @@ static int WriteAlphaPlane(FILE* fout, const WebPDecBuffer* const buffer) {
 // format=PGM: save a grayscale PGM file using the IMC4 layout
 // (http://www.fourcc.org/yuv.php#IMC4). This is a very convenient format for
 // viewing the samples, esp. for odd dimensions.
-// format=YUV: just save the Y/U/V/A planes sequentially without header.
+// format=RAW_YUV: just save the Y/U/V/A planes sequentially without header.
 static int WritePGMOrYUV(FILE* fout, const WebPDecBuffer* const buffer,
                          OutputFileFormat format) {
   const int width = buffer->width;
@@ -426,7 +449,7 @@ static int WritePGMOrYUV(FILE* fout, const WebPDecBuffer* const buffer,
   const WebPYUVABuffer* const yuv = &buffer->u.YUVA;
   int ok = 1;
   int y;
-  const int pad = (format == YUV) ? 0 : 1;
+  const int pad = (format == RAW_YUV) ? 0 : 1;
   const int uv_width = (width + 1) / 2;
   const int uv_height = (height + 1) / 2;
   const int out_stride = (width + pad) & ~pad;
@@ -487,7 +510,9 @@ static int SaveOutput(const WebPDecBuffer* const buffer,
     }
   }
 
-  if (format == PNG) {
+  if (format == PNG ||
+      format == RGBA || format == BGRA || format == ARGB ||
+      format == rgbA || format == bgrA || format == Argb) {
 #ifdef HAVE_WINCODEC_H
     ok &= WritePNG(out_file, use_stdout, buffer);
 #else
@@ -495,14 +520,17 @@ static int SaveOutput(const WebPDecBuffer* const buffer,
 #endif
   } else if (format == PAM) {
     ok &= WritePPM(fout, buffer, 1);
-  } else if (format == PPM) {
+  } else if (format == PPM || format == RGB || format == BGR) {
     ok &= WritePPM(fout, buffer, 0);
+  } else if (format == RGBA_4444 || format == RGB_565 || format == rgbA_4444) {
+    ok &= Write16bAsPGM(fout, buffer);
   } else if (format == BMP) {
     ok &= WriteBMP(fout, buffer);
   } else if (format == TIFF) {
     ok &= WriteTIFF(fout, buffer);
-  } else if (format == PGM || format == YUV) {
+  } else if (format == PGM || format == RAW_YUV ||
-    ok &= WritePGMOrYUV(fout, buffer, format);
+             format == YUV || format == YUVA) {
+    ok &= WritePGMOrYUV(fout, buffer, format == RAW_YUV ? RAW_YUV : PGM);
   } else if (format == ALPHA_PLANE_ONLY) {
     ok &= WriteAlphaPlane(fout, buffer);
   }
@@ -569,6 +597,70 @@ static const char* const kFormatType[] = {
   "unspecified", "lossy", "lossless"
 };
 
+static uint8_t* AllocateExternalBuffer(WebPDecoderConfig* config,
+                                       OutputFileFormat format,
+                                       int use_external_memory) {
+  uint8_t* external_buffer = NULL;
+  WebPDecBuffer* const output_buffer = &config->output;
+  int w = config->input.width;
+  int h = config->input.height;
+  if (config->options.use_scaling) {
+    w = config->options.scaled_width;
+    h = config->options.scaled_height;
+  } else if (config->options.use_cropping) {
+    w = config->options.crop_width;
+    h = config->options.crop_height;
+  }
+  if (format >= RGB && format <= rgbA_4444) {
+    const int bpp = (format == RGB || format == BGR) ? 3
+                  : (format == RGBA_4444 || format == rgbA_4444 ||
+                     format == RGB_565) ? 2
+                  : 4;
+    uint32_t stride = bpp * w + 7;   // <- just for exercising
+    external_buffer = (uint8_t*)malloc(stride * h);
+    if (external_buffer == NULL) return NULL;
+    output_buffer->u.RGBA.stride = stride;
+    output_buffer->u.RGBA.size = stride * h;
+    output_buffer->u.RGBA.rgba = external_buffer;
+  } else {    // YUV and YUVA
+    const int has_alpha = WebPIsAlphaMode(output_buffer->colorspace);
+    uint8_t* tmp;
+    uint32_t stride = w + 3;
+    uint32_t uv_stride = (w + 1) / 2 + 13;
+    uint32_t total_size = stride * h * (has_alpha ? 2 : 1)
+                        + 2 * uv_stride * (h + 1) / 2;
+    assert(format >= YUV && format <= YUVA);
+    external_buffer = (uint8_t*)malloc(total_size);
+    if (external_buffer == NULL) return NULL;
+    tmp = external_buffer;
+    output_buffer->u.YUVA.y = tmp;
+    output_buffer->u.YUVA.y_stride = stride;
+    output_buffer->u.YUVA.y_size = stride * h;
+    tmp += output_buffer->u.YUVA.y_size;
+    if (has_alpha) {
+      output_buffer->u.YUVA.a = tmp;
+      output_buffer->u.YUVA.a_stride = stride;
+      output_buffer->u.YUVA.a_size = stride * h;
+      tmp += output_buffer->u.YUVA.a_size;
+    } else {
+      output_buffer->u.YUVA.a = NULL;
+      output_buffer->u.YUVA.a_stride = 0;
+    }
+    output_buffer->u.YUVA.u = tmp;
+    output_buffer->u.YUVA.u_stride = uv_stride;
+    output_buffer->u.YUVA.u_size = uv_stride * (h + 1) / 2;
+    tmp += output_buffer->u.YUVA.u_size;
+
+    output_buffer->u.YUVA.v = tmp;
+    output_buffer->u.YUVA.v_stride = uv_stride;
+    output_buffer->u.YUVA.v_size = uv_stride * (h + 1) / 2;
+    tmp += output_buffer->u.YUVA.v_size;
+    assert(tmp <= external_buffer + total_size);
+  }
+  output_buffer->is_external_memory = use_external_memory;
+  return external_buffer;
+}
+
 int main(int argc, const char *argv[]) {
   int ok = 0;
   const char *in_file = NULL;
@@ -578,6 +670,10 @@ int main(int argc, const char *argv[]) {
   WebPDecBuffer* const output_buffer = &config.output;
   WebPBitstreamFeatures* const bitstream = &config.input;
   OutputFileFormat format = PNG;
+  uint8_t* external_buffer = NULL;
+  int use_external_memory = 0;
+  const uint8_t* data = NULL;
+
   int incremental = 0;
   int c;
 
@@ -617,7 +713,32 @@ int main(int argc, const char *argv[]) {
     } else if (!strcmp(argv[c], "-pgm")) {
       format = PGM;
     } else if (!strcmp(argv[c], "-yuv")) {
-      format = YUV;
+      format = RAW_YUV;
+    } else if (!strcmp(argv[c], "-pixel_format") && c < argc - 1) {
+      const char* const fmt = argv[++c];
+      if      (!strcmp(fmt, "RGB"))  format = RGB;
+      else if (!strcmp(fmt, "RGBA")) format = RGBA;
+      else if (!strcmp(fmt, "BGR"))  format = BGR;
+      else if (!strcmp(fmt, "BGRA")) format = BGRA;
+      else if (!strcmp(fmt, "ARGB")) format = ARGB;
+      else if (!strcmp(fmt, "RGBA_4444")) format = RGBA_4444;
+      else if (!strcmp(fmt, "RGB_565")) format = RGB_565;
+      else if (!strcmp(fmt, "rgbA")) format = rgbA;
+      else if (!strcmp(fmt, "bgrA")) format = bgrA;
+      else if (!strcmp(fmt, "Argb")) format = Argb;
+      else if (!strcmp(fmt, "rgbA_4444")) format = rgbA_4444;
+      else if (!strcmp(fmt, "YUV"))  format = YUV;
+      else if (!strcmp(fmt, "YUVA")) format = YUVA;
+      else {
+        fprintf(stderr, "Can't parse pixel_format %s\n", fmt);
+        parse_error = 1;
+      }
+    } else if (!strcmp(argv[c], "-external_memory") && c < argc - 1) {
+      use_external_memory = ExUtilGetInt(argv[++c], 0, &parse_error);
+      parse_error |= (use_external_memory > 2 || use_external_memory < 0);
+      if (parse_error) {
+        fprintf(stderr, "Can't parse 'external_memory' value %s\n", argv[c]);
+      }
     } else if (!strcmp(argv[c], "-mt")) {
       config.options.use_threads = 1;
     } else if (!strcmp(argv[c], "-alpha_dither")) {
@@ -676,7 +797,6 @@ int main(int argc, const char *argv[]) {
   {
     VP8StatusCode status = VP8_STATUS_OK;
     size_t data_size = 0;
-    const uint8_t* data = NULL;
     if (!ExUtilLoadWebP(in_file, &data, &data_size, bitstream)) {
       return -1;
     }
@@ -703,15 +823,33 @@ int main(int argc, const char *argv[]) {
             bitstream->has_alpha ? MODE_rgbA : MODE_RGB;
         break;
       case PGM:
-      case YUV:
+      case RAW_YUV:
         output_buffer->colorspace = bitstream->has_alpha ? MODE_YUVA : MODE_YUV;
         break;
       case ALPHA_PLANE_ONLY:
         output_buffer->colorspace = MODE_YUVA;
         break;
-      default:
+      // forced modes:
-        free((void*)data);
+      case RGB: output_buffer->colorspace = MODE_RGB; break;
-        return -1;
+      case RGBA: output_buffer->colorspace = MODE_RGBA; break;
+      case BGR: output_buffer->colorspace = MODE_BGR; break;
+      case BGRA: output_buffer->colorspace = MODE_BGRA; break;
+      case ARGB: output_buffer->colorspace = MODE_ARGB; break;
+      case RGBA_4444: output_buffer->colorspace = MODE_RGBA_4444; break;
+      case RGB_565: output_buffer->colorspace = MODE_RGB_565; break;
+      case rgbA: output_buffer->colorspace = MODE_rgbA; break;
+      case bgrA: output_buffer->colorspace = MODE_bgrA; break;
+      case Argb: output_buffer->colorspace = MODE_Argb; break;
+      case rgbA_4444: output_buffer->colorspace = MODE_rgbA_4444; break;
+      case YUV: output_buffer->colorspace = MODE_YUV; break;
+      case YUVA: output_buffer->colorspace = MODE_YUVA; break;
+      default: goto Exit;
+    }
+
+    if (use_external_memory > 0 && format >= RGB) {
+      external_buffer = AllocateExternalBuffer(&config, format,
+                                               use_external_memory);
+      if (external_buffer == NULL) goto Exit;
     }
 
     if (incremental) {
@@ -720,7 +858,6 @@ int main(int argc, const char *argv[]) {
       status = ExUtilDecodeWebP(data, data_size, verbose, &config);
     }
 
-    free((void*)data);
     ok = (status == VP8_STATUS_OK);
     if (!ok) {
       ExUtilPrintWebPError(in_file, status);
@@ -750,6 +887,8 @@ int main(int argc, const char *argv[]) {
   }
  Exit:
   WebPFreeDecBuffer(output_buffer);
+  free((void*)external_buffer);
+  free((void*)data);
   return ok ? 0 : -1;
 }
 

src/dec/buffer.c

@@ -92,7 +92,7 @@ static VP8StatusCode AllocateBuffer(WebPDecBuffer* const buffer) {
     return VP8_STATUS_INVALID_PARAM;
   }
 
-  if (!buffer->is_external_memory && buffer->private_memory == NULL) {
+  if (buffer->is_external_memory <= 0 && buffer->private_memory == NULL) {
     uint8_t* output;
     int uv_stride = 0, a_stride = 0;
     uint64_t uv_size = 0, a_size = 0, total_size;
@@ -227,7 +227,7 @@ int WebPInitDecBufferInternal(WebPDecBuffer* buffer, int version) {
 
 void WebPFreeDecBuffer(WebPDecBuffer* buffer) {
   if (buffer != NULL) {
-    if (!buffer->is_external_memory) {
+    if (buffer->is_external_memory <= 0) {
       WebPSafeFree(buffer->private_memory);
     }
     buffer->private_memory = NULL;
@@ -256,5 +256,45 @@ void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst) {
   }
 }
 
-//------------------------------------------------------------------------------
+VP8StatusCode WebPCopyDecBufferPixels(const WebPDecBuffer* const src_buf,
+                                      WebPDecBuffer* const dst_buf) {
+  assert(src_buf != NULL && dst_buf != NULL);
+  assert(src_buf->colorspace == dst_buf->colorspace);
 
+  dst_buf->width = src_buf->width;
+  dst_buf->height = src_buf->height;
+  if (CheckDecBuffer(dst_buf) != VP8_STATUS_OK) {
+    return VP8_STATUS_INVALID_PARAM;
+  }
+  if (WebPIsRGBMode(src_buf->colorspace)) {
+    const WebPRGBABuffer* const src = &src_buf->u.RGBA;
+    const WebPRGBABuffer* const dst = &dst_buf->u.RGBA;
+    WebPCopyPlane(src->rgba, src->stride, dst->rgba, dst->stride,
+                  src_buf->width * kModeBpp[src_buf->colorspace],
+                  src_buf->height);
+  } else {
+    const WebPYUVABuffer* const src = &src_buf->u.YUVA;
+    const WebPYUVABuffer* const dst = &dst_buf->u.YUVA;
+    WebPCopyPlane(src->y, src->y_stride, dst->y, dst->y_stride,
+                  src_buf->width, src_buf->height);
+    WebPCopyPlane(src->u, src->u_stride, dst->u, dst->u_stride,
+                  (src_buf->width + 1) / 2, (src_buf->height + 1) / 2);
+    WebPCopyPlane(src->v, src->v_stride, dst->v, dst->v_stride,
+                  (src_buf->width + 1) / 2, (src_buf->height + 1) / 2);
+    if (WebPIsAlphaMode(src_buf->colorspace)) {
+      WebPCopyPlane(src->a, src->a_stride, dst->a, dst->a_stride,
+                    src_buf->width, src_buf->height);
+    }
+  }
+  return VP8_STATUS_OK;
+}
+
+int WebPAvoidSlowMemory(const WebPDecBuffer* const output,
+                        const WebPBitstreamFeatures* const features) {
+  assert(output != NULL);
+  return (output->is_external_memory >= 2) &&
+         WebPIsPremultipliedMode(output->colorspace) &&
+         (features != NULL && features->has_alpha);
+}
+
+//------------------------------------------------------------------------------

src/dec/idec.c

@@ -70,7 +70,9 @@ struct WebPIDecoder {
   VP8Io io_;
 
   MemBuffer mem_;          // input memory buffer.
-  WebPDecBuffer output_;   // output buffer (when no external one is supplied)
+  WebPDecBuffer output_;   // output buffer (when no external one is supplied,
+                           // or if the external one has slow-memory)
+  WebPDecBuffer* final_output_;  // Slow-memory output to copy to eventually.
   size_t chunk_size_;      // Compressed VP8/VP8L size extracted from Header.
 
   int last_mb_y_;          // last row reached for intra-mode decoding
@@ -249,10 +251,16 @@ static VP8StatusCode FinishDecoding(WebPIDecoder* const idec) {
 
   idec->state_ = STATE_DONE;
   if (options != NULL && options->flip) {
-    return WebPFlipBuffer(output);
+    const VP8StatusCode status = WebPFlipBuffer(output);
-  } else {
+    if (status != VP8_STATUS_OK) return status;
-    return VP8_STATUS_OK;
   }
+  if (idec->final_output_ != NULL) {
+    WebPCopyDecBufferPixels(output, idec->final_output_);  // do the slow-copy
+    WebPFreeDecBuffer(&idec->output_);
+    *output = *idec->final_output_;
+    idec->final_output_ = NULL;
+  }
+  return VP8_STATUS_OK;
 }
 
 //------------------------------------------------------------------------------
@@ -575,9 +583,10 @@ static VP8StatusCode IDecode(WebPIDecoder* idec) {
 }
 
 //------------------------------------------------------------------------------
-// Public functions
+// Internal constructor
 
-WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
+static WebPIDecoder* NewDecoder(WebPDecBuffer* const output_buffer,
+                                const WebPBitstreamFeatures* const features) {
   WebPIDecoder* idec = (WebPIDecoder*)WebPSafeCalloc(1ULL, sizeof(*idec));
   if (idec == NULL) {
     return NULL;
@@ -593,25 +602,46 @@ WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
   VP8InitIo(&idec->io_);
 
   WebPResetDecParams(&idec->params_);
-  idec->params_.output = (output_buffer != NULL) ? output_buffer
+  if (output_buffer == NULL || WebPAvoidSlowMemory(output_buffer, features)) {
-                                                 : &idec->output_;
+    idec->params_.output = &idec->output_;
+    idec->final_output_ = output_buffer;
+    if (output_buffer != NULL) {
+      idec->params_.output->colorspace = output_buffer->colorspace;
+    }
+  } else {
+    idec->params_.output = output_buffer;
+    idec->final_output_ = NULL;
+  }
   WebPInitCustomIo(&idec->params_, &idec->io_);  // Plug the I/O functions.
 
   return idec;
 }
 
+//------------------------------------------------------------------------------
+// Public functions
+
+WebPIDecoder* WebPINewDecoder(WebPDecBuffer* output_buffer) {
+  return NewDecoder(output_buffer, NULL);
+}
+
 WebPIDecoder* WebPIDecode(const uint8_t* data, size_t data_size,
                           WebPDecoderConfig* config) {
   WebPIDecoder* idec;
+  WebPBitstreamFeatures tmp_features;
+  WebPBitstreamFeatures* const features =
+      (config == NULL) ? &tmp_features : &config->input;
+  memset(&tmp_features, 0, sizeof(tmp_features));
 
   // Parse the bitstream's features, if requested:
-  if (data != NULL && data_size > 0 && config != NULL) {
+  if (data != NULL && data_size > 0) {
-    if (WebPGetFeatures(data, data_size, &config->input) != VP8_STATUS_OK) {
+    if (WebPGetFeatures(data, data_size, features) != VP8_STATUS_OK) {
       return NULL;
     }
   }
+
   // Create an instance of the incremental decoder
-  idec = WebPINewDecoder(config ? &config->output : NULL);
+  idec = (config != NULL) ? NewDecoder(&config->output, features)
+                          : NewDecoder(NULL, features);
   if (idec == NULL) {
     return NULL;
   }
@@ -645,11 +675,11 @@ void WebPIDelete(WebPIDecoder* idec) {
 
 WebPIDecoder* WebPINewRGB(WEBP_CSP_MODE mode, uint8_t* output_buffer,
                           size_t output_buffer_size, int output_stride) {
-  const int is_external_memory = (output_buffer != NULL);
+  const int is_external_memory = (output_buffer != NULL) ? 1 : 0;
   WebPIDecoder* idec;
 
   if (mode >= MODE_YUV) return NULL;
-  if (!is_external_memory) {    // Overwrite parameters to sane values.
+  if (is_external_memory == 0) {    // Overwrite parameters to sane values.
     output_buffer_size = 0;
     output_stride = 0;
   } else {  // A buffer was passed. Validate the other params.
@@ -671,11 +701,11 @@ WebPIDecoder* WebPINewYUVA(uint8_t* luma, size_t luma_size, int luma_stride,
                            uint8_t* u, size_t u_size, int u_stride,
                            uint8_t* v, size_t v_size, int v_stride,
                            uint8_t* a, size_t a_size, int a_stride) {
-  const int is_external_memory = (luma != NULL);
+  const int is_external_memory = (luma != NULL) ? 1 : 0;
   WebPIDecoder* idec;
   WEBP_CSP_MODE colorspace;
 
-  if (!is_external_memory) {    // Overwrite parameters to sane values.
+  if (is_external_memory == 0) {    // Overwrite parameters to sane values.
     luma_size = u_size = v_size = a_size = 0;
     luma_stride = u_stride = v_stride = a_stride = 0;
     u = v = a = NULL;
@@ -783,6 +813,9 @@ static const WebPDecBuffer* GetOutputBuffer(const WebPIDecoder* const idec) {
   if (idec->state_ <= STATE_VP8_PARTS0) {
     return NULL;
   }
+  if (idec->final_output_ != NULL) {
+    return NULL;   // not yet slow-copied
+  }
   return idec->params_.output;
 }
 
@@ -792,7 +825,7 @@ const WebPDecBuffer* WebPIDecodedArea(const WebPIDecoder* idec,
   const WebPDecBuffer* const src = GetOutputBuffer(idec);
   if (left != NULL) *left = 0;
   if (top != NULL) *top = 0;
-  if (src) {
+  if (src != NULL) {
     if (width != NULL) *width = src->width;
     if (height != NULL) *height = idec->params_.last_y;
   } else {

src/dec/vp8i.h

@@ -32,7 +32,7 @@ extern "C" {
 // version numbers
 #define DEC_MAJ_VERSION 0
 #define DEC_MIN_VERSION 5
-#define DEC_REV_VERSION 0
+#define DEC_REV_VERSION 1
 
 // YUV-cache parameters. Cache is 32-bytes wide (= one cacheline).
 // Constraints are: We need to store one 16x16 block of luma samples (y),

src/dec/webp.c

@@ -514,6 +514,8 @@ static VP8StatusCode DecodeInto(const uint8_t* const data, size_t data_size,
     WebPFreeDecBuffer(params->output);
   } else {
     if (params->options != NULL && params->options->flip) {
+      // This restores the original stride values if options->flip was used
+      // during the call to WebPAllocateDecBuffer above.
       status = WebPFlipBuffer(params->output);
     }
   }
@@ -758,9 +760,24 @@ VP8StatusCode WebPDecode(const uint8_t* data, size_t data_size,
   }
 
   WebPResetDecParams(&params);
-  params.output = &config->output;
   params.options = &config->options;
-  status = DecodeInto(data, data_size, &params);
+  params.output = &config->output;
+  if (WebPAvoidSlowMemory(params.output, &config->input)) {
+    // decoding to slow memory: use a temporary in-mem buffer to decode into.
+    WebPDecBuffer in_mem_buffer;
+    WebPInitDecBuffer(&in_mem_buffer);
+    in_mem_buffer.colorspace = config->output.colorspace;
+    in_mem_buffer.width = config->input.width;
+    in_mem_buffer.height = config->input.height;
+    params.output = &in_mem_buffer;
+    status = DecodeInto(data, data_size, &params);
+    if (status == VP8_STATUS_OK) {  // do the slow-copy
+      status = WebPCopyDecBufferPixels(&in_mem_buffer, &config->output);
+    }
+    WebPFreeDecBuffer(&in_mem_buffer);
+  } else {
+    status = DecodeInto(data, data_size, &params);
+  }
 
   return status;
 }
@@ -826,4 +843,3 @@ int WebPIoInitFromOptions(const WebPDecoderOptions* const options,
 }
 
 //------------------------------------------------------------------------------
-

src/dec/webpi.h

@@ -45,11 +45,20 @@ struct WebPDecParams {
   OutputFunc emit;               // output RGB or YUV samples
   OutputAlphaFunc emit_alpha;    // output alpha channel
   OutputRowFunc emit_alpha_row;  // output one line of rescaled alpha values
+
+  WebPDecBuffer* final_output;   // In case the user supplied a slow-memory
+                                 // output, we decode image in temporary buffer
+                                 // (this::output) and copy it here.
+  WebPDecBuffer tmp_buffer;      // this::output will point to this one in case
+                                 // of slow memory.
 };
 
 // Should be called first, before any use of the WebPDecParams object.
 void WebPResetDecParams(WebPDecParams* const params);
 
+// Delete all memory (after an error occurred, for instance)
+void WebPFreeDecParams(WebPDecParams* const params);
+
 //------------------------------------------------------------------------------
 // Header parsing helpers
 
@@ -107,13 +116,23 @@ VP8StatusCode WebPAllocateDecBuffer(int width, int height,
 VP8StatusCode WebPFlipBuffer(WebPDecBuffer* const buffer);
 
 // Copy 'src' into 'dst' buffer, making sure 'dst' is not marked as owner of the
-// memory (still held by 'src').
+// memory (still held by 'src'). No pixels are copied.
 void WebPCopyDecBuffer(const WebPDecBuffer* const src,
                        WebPDecBuffer* const dst);
 
 // Copy and transfer ownership from src to dst (beware of parameter order!)
 void WebPGrabDecBuffer(WebPDecBuffer* const src, WebPDecBuffer* const dst);
 
+// Copy pixels from 'src' into a *preallocated* 'dst' buffer. Returns
+// VP8_STATUS_INVALID_PARAM if the 'dst' is not set up correctly for the copy.
+VP8StatusCode WebPCopyDecBufferPixels(const WebPDecBuffer* const src,
+                                      WebPDecBuffer* const dst);
+
+// Returns true if decoding will be slow with the current configuration
+// and bitstream features.
+int WebPAvoidSlowMemory(const WebPDecBuffer* const output,
+                        const WebPBitstreamFeatures* const features);
+
 //------------------------------------------------------------------------------
 
 #ifdef __cplusplus

src/webp/decode.h

@@ -197,7 +197,10 @@ struct WebPYUVABuffer {              // view as YUVA
 struct WebPDecBuffer {
   WEBP_CSP_MODE colorspace;  // Colorspace.
   int width, height;         // Dimensions.
-  int is_external_memory;    // If true, 'internal_memory' pointer is not used.
+  int is_external_memory;    // If non-zero, 'internal_memory' pointer is not
+                             // used. If value is '2' or more, the external
+                             // memory is considered 'slow' and multiple
+                             // read/write will be avoided.
   union {
     WebPRGBABuffer RGBA;
     WebPYUVABuffer YUVA;
@@ -205,7 +208,7 @@ struct WebPDecBuffer {
   uint32_t       pad[4];     // padding for later use
 
   uint8_t* private_memory;   // Internally allocated memory (only when
-                             // is_external_memory is false). Should not be used
+                             // is_external_memory is 0). Should not be used
                              // externally, but accessed via the buffer union.
 };
 
@@ -269,7 +272,7 @@ typedef enum VP8StatusCode {
 // that of the returned WebPIDecoder object.
 // The supplied 'output_buffer' content MUST NOT be changed between calls to
 // WebPIAppend() or WebPIUpdate() unless 'output_buffer.is_external_memory' is
-// set to 1. In such a case, it is allowed to modify the pointers, size and
+// not set to 0. In such a case, it is allowed to modify the pointers, size and
 // stride of output_buffer.u.RGBA or output_buffer.u.YUVA, provided they remain
 // within valid bounds.
 // All other fields of WebPDecBuffer MUST remain constant between calls.