Building a fast PNG encoder issues - compression

Question

I am trying to build a fast 8-bit greyscale PNG encoder. Unfortunately I must be misunderstanding part of the spec. Smaller image sizes seem to work, but the larger ones will only open in some image viewers. This image (with multiple DEFLATE Blocks) gives a
"Decompression error in IDAT" error in my image viewer but opens fine in my browser:
This image has just one DEFLATE block but also gives an error:
Below I will outline what I put in my IDAT chunk in case you can easily spot any mistakes (note, images and steps have been modified based on answers, but there is still a problem):
IDAT length
"IDAT" in ascii (literally the bytes 0x49 0x44 0x41 0x54)
Zlib header 0x78 0x01
Steps 4-7 are for every deflate block, as the data may need to be broken up:
The byte 0x00 or 0x01, depending on if it is a middle or the last block.
Number of bytes in block (up to 2^16-1) stored as a little endian 16-bit integer
The 1's complement of this integer representation.
Image data (each scan-line is starts with a zero-byte for the no filter option in PNG, and is followed by width bytes of greyscale pixel data)
An adler-32 checksum of all the image data
A CRC of all the IDAT data
I've tried pngcheck on linux, an it does not spot any errors. If nobody can see what is wrong, can you point me in the right direction for a debugging tool?
My last resort is to use the libpng library to make my own decoder, and debug from there.
Some people have suggested it may be my adler-32 function calculation:
static uint32_t adler32(uint32_t height, uint32_t width, char** pixel_array)
{
uint32_t a=1,b=0,w,h;
for(h=0;h<height;h++)
{
b+=a;
for(w=0;w<width;w++)
{
a+=pixel_array[h][w];
b+=a;
}
}
return (uint32_t)(((b%65521)*65536)|(a%65521));
}
Note that because the pixel_array passed to the function does not contain the zero-byte at the beginning of each scanline (needed for PNG) there is an extra b+=a (and implicit a+=0) at the beginning of each iteration of the outer loop.

Answer 1

I do get an error with pngcheck: "zlib: inflate error = -3 (data error)". As your PNG scaffolding structure looks okay, it's time to take a low-level look into the IDAT block with a hex viewer. (I'm going to type this up while working through it.)
The header looks alright; IDAT length is okay. Your zlib flags are 78 01 ("No/low compression", see also What does a zlib header look like?), where one of my own tools use 78 9C ("Default compression"), but then again, these flags are only informative.
Next: zlib's internal blocks (per RFC1950).
Directly after the compression flags (CMF in RFC1950) it expects FLATE compressed data, which is the only compression scheme zlib supports. And that is in another castle RFC: RFC1951.
Each separately compression block is prepended by a byte:
3.2.3. Details of block format
Each block of compressed data begins with 3 header bits
containing the following data:
first bit BFINAL
next 2 bits BTYPE
...
BFINAL is set if and only if this is the last block of the data
set.
BTYPE specifies how the data are compressed, as follows:
00 - no compression
01 - compressed with fixed Huffman codes
10 - compressed with dynamic Huffman codes
11 - reserved (error)
So this value can be set to 00 for 'not last block, uncompressed' and to 01 for 'last block, uncompressed', immediately followed by the length (2 bytes) and its bitwise inverse, per 3.2.4. Non-compressed blocks (BTYPE=00):
3.2.4. Non-compressed blocks (BTYPE=00)
Any bits of input up to the next byte boundary are ignored.
The rest of the block consists of the following information:
0 1 2 3 4...
+---+---+---+---+================================+
| LEN | NLEN |... LEN bytes of literal data...|
+---+---+---+---+================================+
LEN is the number of data bytes in the block. NLEN is the
one's complement of LEN.
They are the final 4 bytes in your IDAT segment. Why do small images work, and larger not? Because you only have 2 bytes for the length.1 You need to break up your image into blocks no larger than 65,535 bytes (in my own PNG creator I seem to have used 32,768, probably "for safety"). If the last block, write out 01, else 00. Then add the two times two LEN bytes, properly encoded, followed by exactly LEN data bytes. Repeat until done.
The Adler-32 checksum is not part of this Flate-compressed data, and should not be counted in the blocks of LEN data. (It is still part of the IDAT block, though.)
After re-reading your question to verify I addressed all of your issues (and confirming I spelled "Adler-32" correctly), I realized you describe all of the steps right -- except that the 'last block' indicator is 01, not 80 (later edit: uh, perhaps you are right about that!) -- but that it does not show in this sample PNG. See if you can get it to work following all of the steps by the letter.
Kudos for doing this 'by hand'. It's a nice exercise in 'following the specs', and if you get this to work, it may be worthwhile to try and add proper compression. I shun pre-made libraries as much as possible; the only allowance I made for my own PNG encoder/decoder was to use Rich Geldreich's miniz.c, because implementing proper Flate encoding/decoding is beyond my ken.
1 That's not the whole story. Browsers are particularly forgiving in HTML errors; it seems they are as forgiving for PNG errors as well. Safari displays your image just fine, and so does Preview. But they may just all be sharing OS X's PNG decoder, because Photoshop rejects the file.

Answer 2

The byte 0x00 or 0x80, depending on if it is a middle or the last block.
Change the 0x80 to 0x01 and all will be well.
The 0x80 is appearing as a stored block that is not the last block. All that's being looked at is the low bit, which is zero, indicating a middle block. All of the data is in that "middle" block, so a decoder will recover the full image. Some liberal PNG decoders may then ignore the errors it gets when it tries to decode the next block, which isn't there, and then ignore the missing check values (Adler-32 and CRC-32), etc. That's why it shows up ok in browsers, even though it is an invalid PNG file.
There are two things wrong with your Adler-32 code. First, you are accessing the data from a char array. char is signed, so your 0xff bytes are being added not as 255, but rather as -127. You need to make the array unsigned char or cast it to that before extracting byte values from it.
Second, you are doing the modulo operation too late. You must do the % 65521 before the uint32_t overflows. Otherwise you don't get the modulo of the sum as required by the algorithm. A simple fix would be to do the % 65521 to a and b right after the width loop, inside the height loop. This will work so long as you can guarantee that the width will be less than 5551 bytes. (Why 5551 is left as an exercise for the reader.) If you cannot guarantee that, then you will need to embed a another loop to consume bytes from the line until you get to 5551 of them, do the modulo, and then continue with the line. Or, a smidge slower, just run a counter and do the modulo when it gets to the limit.
Here is an example of a version that works for any width:
static uint32_t adler32(uint32_t height, uint32_t width, unsigned char ** pixel_array)
{
uint32_t a = 1, b = 0, w, h, k;
for (h = 0; h < height; h++)
{
b += a;
w = k = 0;
while (k < width) {
k += 5551;
if (k > width)
k = width;
while (w < k) {
a += pixel_array[h][w++];
b += a;
}
a %= 65521;
b %= 65521;
}
}
return (b << 16) | a;
}