Merge pull request #4181 from nnethercote/compact-images

Reduce memory consumption of simple black and white images.

src/core/evaluator.js

@@ -166,7 +166,7 @@ var PartialEvaluator = (function PartialEvaluatorClosure() {
           (w + h) < SMALL_IMAGE_DIMENSIONS) {
         var imageObj = new PDFImage(this.xref, resources, image,
                                     inline, null, null);
-        var imgData = imageObj.getImageData();
+        var imgData = imageObj.createImageData();
         operatorList.addOp(OPS.paintInlineImageXObject, [imgData]);
         return;
       }
@@ -189,7 +189,7 @@ var PartialEvaluator = (function PartialEvaluatorClosure() {
 
 
       PDFImage.buildImage(function(imageObj) {
-          var imgData = imageObj.getImageData();
+          var imgData = imageObj.createImageData();
           self.handler.send('obj', [objId, self.pageIndex, 'Image', imgData],
                             null, [imgData.data.buffer]);
         }, self.handler, self.xref, resources, image, inline);
@@ -1318,7 +1318,8 @@ var PartialEvaluator = (function PartialEvaluatorClosure() {
         // replacing queue items
         squash(fnArray, j, count * 4, OPS.paintInlineImageXObjectGroup);
         argsArray.splice(j, count * 4,
-          [{width: imgWidth, height: imgHeight, data: imgData}, map]);
+          [{width: imgWidth, height: imgHeight, kind: 'rgba_32bpp',
+            data: imgData}, map]);
         i = j;
         ii = argsArray.length;
       }

src/core/image.js

@@ -267,10 +267,10 @@ var PDFImage = (function PDFImageClosure() {
       var bpc = this.bpc;
 
       // This image doesn't require any extra work.
-      if (bpc === 8)
+      if (bpc === 8) {
         return buffer;
+      }
 
-      var bufferLength = buffer.length;
       var width = this.width;
       var height = this.height;
       var numComps = this.numComps;
@@ -417,7 +417,14 @@ var PDFImage = (function PDFImageClosure() {
         buffer[i + 2] = clamp((buffer[i + 2] - matteRgb[2]) * k + matteRgb[2]);
       }
     },
-    fillRgbaBuffer: function PDFImage_fillRgbaBuffer(buffer, width, height) {
+    createImageData: function PDFImage_createImageData() {
+      var drawWidth = this.drawWidth;
+      var drawHeight = this.drawHeight;
+      var imgData = {       // other fields are filled in below
+        width: drawWidth,
+        height: drawHeight,
+      };
+
       var numComps = this.numComps;
       var originalWidth = this.width;
       var originalHeight = this.height;
@@ -429,21 +436,46 @@ var PDFImage = (function PDFImageClosure() {
 
       // imgArray can be incomplete (e.g. after CCITT fax encoding)
       var actualHeight = 0 | (imgArray.length / rowBytes *
-                         height / originalHeight);
+                         drawHeight / originalHeight);
+
+      // If it is a 1-bit-per-pixel grayscale (i.e. black-and-white) image
+      // without any complications, we pass a same-sized copy to the main
+      // thread rather than expanding by 32x to RGBA form. This saves *lots* of
+      // memory for many scanned documents. It's also much faster.
+      if (this.colorSpace.name === 'DeviceGray' && bpc === 1 &&
+          !this.smask && !this.mask && !this.needsDecode &&
+          drawWidth === originalWidth && drawHeight === originalHeight) {
+        imgData.kind = 'grayscale_1bpp';
+
+        // We must make a copy of imgArray, otherwise it'll be neutered upon
+        // transfer which will break any code that subsequently reuses it.
+        var newArray = new Uint8Array(imgArray.length);
+        newArray.set(imgArray);
+        imgData.data = newArray;
+        imgData.origLength = imgArray.length;
+        return imgData;
+      }
+
       var comps = this.getComponents(imgArray);
 
+      var rgbaBuf = new Uint8Array(drawWidth * drawHeight * 4);
+
       // Handle opacity here since color key masking needs to be performed on
       // undecoded values.
-      this.fillOpacity(buffer, width, height, actualHeight, comps);
+      this.fillOpacity(rgbaBuf, drawWidth, drawHeight, actualHeight, comps);
 
       if (this.needsDecode) {
         this.decodeBuffer(comps);
       }
 
-      this.colorSpace.fillRgb(buffer, originalWidth, originalHeight, width,
+      this.colorSpace.fillRgb(rgbaBuf, originalWidth, originalHeight, drawWidth,
-                              height, actualHeight, bpc, comps);
+                              drawHeight, actualHeight, bpc, comps);
+
+      this.undoPreblend(rgbaBuf, drawWidth, actualHeight);
 
-      this.undoPreblend(buffer, width, actualHeight);
+      imgData.kind = 'rgba_32bpp';
+      imgData.data = rgbaBuf;
+      return imgData;
     },
     fillGrayBuffer: function PDFImage_fillGrayBuffer(buffer) {
       var numComps = this.numComps;
@@ -468,18 +500,6 @@ var PDFImage = (function PDFImageClosure() {
       for (var i = 0; i < length; ++i)
         buffer[i] = (scale * comps[i]) | 0;
     },
-    getImageData: function PDFImage_getImageData() {
-      var drawWidth = this.drawWidth;
-      var drawHeight = this.drawHeight;
-      var imgData = {
-        width: drawWidth,
-        height: drawHeight,
-        data: new Uint8Array(drawWidth * drawHeight * 4)
-      };
-      var pixels = imgData.data;
-      this.fillRgbaBuffer(pixels, drawWidth, drawHeight);
-      return imgData;
-    },
     getImageBytes: function PDFImage_getImageBytes(length) {
       this.image.reset();
       return this.image.getBytes(length);

src/display/canvas.js

@@ -437,45 +437,81 @@ var CanvasGraphics = (function CanvasGraphicsClosure() {
     // of putImageData(). (E.g. in Firefox we make two short-lived copies of
     // the data passed to putImageData()). |n| shouldn't be too small, however,
     // because too many putImageData() calls will slow things down.
+    //
+    // Note: as written, if the last chunk is partial, the putImageData() call
+    // will (conceptually) put pixels past the bounds of the canvas.  But
+    // that's ok; any such pixels are ignored.
 
-    var rowsInFullChunks = 16;
+    var fullChunkHeight = 16;
-    var fullChunks = (imgData.height / rowsInFullChunks) | 0;
+    var fracChunks = imgData.height / fullChunkHeight;
-    var rowsInLastChunk = imgData.height - fullChunks * rowsInFullChunks;
+    var fullChunks = Math.floor(fracChunks);
-    var elemsInFullChunks = imgData.width * rowsInFullChunks * 4;
+    var totalChunks = Math.ceil(fracChunks);
-    var elemsInLastChunk = imgData.width * rowsInLastChunk * 4;
+    var partialChunkHeight = imgData.height - fullChunks * fullChunkHeight;
 
-    var chunkImgData = ctx.createImageData(imgData.width, rowsInFullChunks);
+    var chunkImgData = ctx.createImageData(imgData.width, fullChunkHeight);
     var srcPos = 0;
     var src = imgData.data;
     var dst = chunkImgData.data;
-    var haveSetAndSubarray = 'set' in dst && 'subarray' in src;
 
-    // Do all the full-size chunks.
+    // There are multiple forms in which the pixel data can be passed, and
-    for (var i = 0; i < fullChunks; i++) {
+    // imgData.kind tells us which one this is.
-      if (haveSetAndSubarray) {
+
-        dst.set(src.subarray(srcPos, srcPos + elemsInFullChunks));
+    if (imgData.kind === 'grayscale_1bpp') {
-        srcPos += elemsInFullChunks;
+      // Grayscale, 1 bit per pixel (i.e. black-and-white).
-      } else {
+      var srcData = imgData.data;
-        for (var j = 0; j < elemsInFullChunks; j++) {
+      var destData = chunkImgData.data;
-          chunkImgData.data[j] = imgData.data[srcPos++];
+      var alpha = 255;
+      var origLength = imgData.origLength;
+      for (var i = 0; i < totalChunks; i++) {
+        var thisChunkHeight =
+          (i < fullChunks) ? fullChunkHeight : partialChunkHeight;
+        var destPos = 0;
+        for (var j = 0; j < thisChunkHeight; j++) {
+          var mask = 0;
+          var srcByte = 0;
+          for (var k = 0; k < imgData.width; k++) {
+            if (srcPos >= origLength) {
+              // We ran out of input. Make all remaining pixels transparent.
+              alpha = 0;
+            }
+            if (mask === 0) {
+              srcByte = srcData[srcPos++];
+              mask = 128;
+            }
+
+            var c = (+!!(srcByte & mask)) * 255;
+            destData[destPos++] = c;
+            destData[destPos++] = c;
+            destData[destPos++] = c;
+            destData[destPos++] = alpha;
+
+            mask >>= 1;
           }
         }
-      ctx.putImageData(chunkImgData, 0, i * rowsInFullChunks);
+        ctx.putImageData(chunkImgData, 0, i * fullChunkHeight);
       }
 
-    // Do the final, partial chunk, if required.
+    } else if (imgData.kind === 'rgba_32bpp') {
-    if (rowsInLastChunk !== 0) {
+      // RGBA, 32-bits per pixel.
+      var haveSetAndSubarray = 'set' in dst && 'subarray' in src;
+
+      for (var i = 0; i < totalChunks; i++) {
+        var thisChunkHeight =
+          (i < fullChunks) ? fullChunkHeight : partialChunkHeight;
+        var elemsInThisChunk = imgData.width * thisChunkHeight * 4;
         if (haveSetAndSubarray) {
-        dst.set(src.subarray(srcPos, srcPos + elemsInLastChunk));
+          dst.set(src.subarray(srcPos, srcPos + elemsInThisChunk));
-        srcPos += elemsInLastChunk;
+          srcPos += elemsInThisChunk;
         } else {
-        for (var j = 0; j < elemsInLastChunk; j++) {
+          for (var j = 0; j < elemsInThisChunk; j++) {
             chunkImgData.data[j] = imgData.data[srcPos++];
           }
         }
-      // This (conceptually) puts pixels past the bounds of the canvas.  But
+        ctx.putImageData(chunkImgData, 0, i * fullChunkHeight);
-      // that's ok; any such pixels are ignored.
+      }
-      ctx.putImageData(chunkImgData, 0, fullChunks * rowsInFullChunks);
+
+    } else {
+        error('bad image kind: ' + imgData.kind);
     }
   }