The output resulted from no-compressed png encoder of demo, even though can be display

<a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/us

No-compressed png encoder fails on Firefox about libjxl HOT 6 OPEN

nnphongphu commented on June 13, 2024

No-compressed png encoder fails on Firefox

from libjxl.

Comments (6)

mo271 commented on June 13, 2024

Thanks for bringing this up. Could you point to the file where something fails?!

from libjxl.

nnphongphu commented on June 13, 2024

Thanks for bringing this up. Could you point to the file where something fails?!

libjxl/tools/wasm_demo/no_png.cc

Lines 168 to 215 in b1dc4ff

 // IDAT 

 WriteU32BE(dst, idat_size); 

 WriteU32(dst, 0x54414449); 

 size_t offset = 0; 

 size_t bytes_to_next_row = 0; 

 uint32_t adler_s1 = 1; 

 uint32_t adler_s2 = 0; 

 WriteU8(dst, 0x08); // CM = 8 (deflate), CINFO = 0 (window size = 2**(0+8)) 

 WriteU8(dst, 29); // FCHECK; (FCHECK + 256* CMF) % 31 = 0 

 for (size_t i = 0; i < num_deflate_blocks; ++i) { 

 size_t block_size = data_size - offset; 

 if (block_size > kMaxDeflateBlock) { 

 block_size = kMaxDeflateBlock; 

 } 

 bool is_last = ((i + 1) == num_deflate_blocks); 

 WriteU8(dst, is_last); // btype = 00 (uncompressed) 

 offset += block_size; 

 WriteU16(dst, block_size); 

 WriteU16(dst, ~block_size); 

 while (block_size > 0) { 

 if (bytes_to_next_row == 0) { 

 WriteU8(dst, 0); // filter: raw 

 adler_s2 += adler_s1; 

 bytes_to_next_row = row_size; 

 block_size--; 

 continue; 

 } 

 size_t bytes_to_copy = std::min(block_size, bytes_to_next_row); 

 AdlerCopy(input, dst, bytes_to_copy, &adler_s1, &adler_s2); 

 dst += bytes_to_copy; 

 input += bytes_to_copy; 

 block_size -= bytes_to_copy; 

 bytes_to_next_row -= bytes_to_copy; 

 } 

 } 

 // Fake Adler works well in Chrome; so let's not waste CPU cycles. 

 uint32_t adler = 0; // (adler_s2 << 8) | adler_s1; 

 WriteU32BE(dst, adler); 

 WriteU32BE(dst, 0); // Fake CRC32 

 // IEND 

 WriteU32BE(dst, 0); 

 chunk_start = dst; 

 WriteU32(dst, 0x444E4549); 

 // TODO(eustas): this is fixed value; precalculate? 

 crc32 = CalculateCrc32(chunk_start, dst); 

 WriteU32BE(dst, crc32);

I think it's somewhere here. I checked the output png file with pngcheck and the output is something like this:

File: image.png (4917485 bytes)
chunk IHDR at offset 0x0000c, length 13
1280 x 1280 image, 24-bit RGB, non-interlaced
chunk iCCP at offset 0x00025, length 550
profile name = 1, compression method = 0 (deflate)
compressed profile = 547 bytes
chunk IDAT at offset 0x00257, length 4916866
zlib: deflated, 256-byte window, superfast compression
CRC error in chunk IDAT (computed 33609e4c, expected 00000000)
ERRORS DETECTED in image.png

from libjxl.

nnphongphu commented on June 13, 2024

It's definitely because of the Fake CRC32 at the end of IDAT. Fixed it. But would it cost much performance-wise?

from libjxl.

mo271 commented on June 13, 2024

@eustas probably can give you an answer

Can this be closed?

from libjxl.

eustas commented on June 13, 2024

IIRC we had that checksum calculation in the past, but removed it for the sake of performance.
I can re-add it as a build option. That way "careless" browsers will be able to get max performance, while others will not fail.
Ping me regularly to make that happen, please.

from libjxl.

jonsneyers commented on June 13, 2024

It would be more useful to make this a runtime thing, and have some detection code to test whether correct CRCs are needed or not, computing them only when needed. There's no point in a wasm polyfill that doesn't work on some browsers...

from libjxl.

No-compressed png encoder fails on Firefox about libjxl HOT 6 OPEN

Comments (6)

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	// IDAT
	WriteU32BE(dst, idat_size);
	WriteU32(dst, 0x54414449);
	size_t offset = 0;
	size_t bytes_to_next_row = 0;
	uint32_t adler_s1 = 1;
	uint32_t adler_s2 = 0;
	WriteU8(dst, 0x08); // CM = 8 (deflate), CINFO = 0 (window size = 2**(0+8))
	WriteU8(dst, 29); // FCHECK; (FCHECK + 256* CMF) % 31 = 0
	for (size_t i = 0; i < num_deflate_blocks; ++i) {
	size_t block_size = data_size - offset;
	if (block_size > kMaxDeflateBlock) {
	block_size = kMaxDeflateBlock;
	}
	bool is_last = ((i + 1) == num_deflate_blocks);
	WriteU8(dst, is_last); // btype = 00 (uncompressed)
	offset += block_size;

	WriteU16(dst, block_size);
	WriteU16(dst, ~block_size);
	while (block_size > 0) {
	if (bytes_to_next_row == 0) {
	WriteU8(dst, 0); // filter: raw
	adler_s2 += adler_s1;
	bytes_to_next_row = row_size;
	block_size--;
	continue;
	}
	size_t bytes_to_copy = std::min(block_size, bytes_to_next_row);
	AdlerCopy(input, dst, bytes_to_copy, &adler_s1, &adler_s2);
	dst += bytes_to_copy;
	input += bytes_to_copy;
	block_size -= bytes_to_copy;
	bytes_to_next_row -= bytes_to_copy;
	}
	}
	// Fake Adler works well in Chrome; so let's not waste CPU cycles.
	uint32_t adler = 0; // (adler_s2 << 8) \| adler_s1;
	WriteU32BE(dst, adler);
	WriteU32BE(dst, 0); // Fake CRC32

	// IEND
	WriteU32BE(dst, 0);
	chunk_start = dst;
	WriteU32(dst, 0x444E4549);
	// TODO(eustas): this is fixed value; precalculate?
	crc32 = CalculateCrc32(chunk_start, dst);
	WriteU32BE(dst, crc32);