Commit Graph
140 Commits
Author SHA1 Message Date
Nigel Tao 7b7dca4767 Start a flate decoder.decode_dynamic method. 2017-07-14 16:36:51 +10:00
Nigel Tao 368bd5e4f7 Add a read_u32le method. 2017-07-13 16:20:59 +10:00
Nigel Tao 7e9b3aab67 Add a high_bits method. 2017-07-13 13:23:01 +10:00
Nigel Tao 73a4f2b263 Remove the unnecessary PUFFS_LOW_BITS macro. 2017-07-13 11:46:26 +10:00
Nigel Tao 4302619eba Add "PUFFS" to "ERROR_BAD_PUFFS_VERSION". 2017-07-13 11:34:36 +10:00
Nigel Tao a3302ff1e5 Fix suspending copy_from32. 2017-07-13 11:30:16 +10:00
Nigel Tao e7873b5ca4 Use "coroutine suspension point" consistently. 2017-07-13 10:46:05 +10:00
Nigel Tao d83cf7f3d5 Track copy_from32 bytes remaining after suspend. 2017-07-13 10:30:39 +10:00
Nigel Tao 8c5b50ef1f Let copy_from32 return SHORT_WRITE or SHORT_READ. 2017-07-13 10:14:30 +10:00
Nigel Tao 940417ec6a Don't make previous non-error statuses sticky. 2017-07-13 09:59:22 +10:00
Nigel Tao 7a41ee4da1 Rename the "status" keyword to "suspension". 2017-07-07 16:54:19 +10:00
Nigel Tao a29dcb92db Fix some error messages for "puffs bench". 2017-07-07 15:12:27 +10:00
Nigel Tao 6a34375cb3 Decode flate stored blocks. 2017-07-07 15:08:40 +10:00
Nigel Tao 727073ac89 Make status code names ALL_CAPS; they're macros. 2017-07-06 12:42:09 +10:00
Nigel Tao f7f57eaa21 Tweak the status code int32_t representation. 2017-07-06 12:27:55 +10:00
Nigel Tao 17b518376f Add a packageid keyword. 2017-07-06 11:17:15 +10:00
Nigel Tao 198dd1cab0 Make the __builtin_expect argument a bool.
No significant benchmark numbers changed. All existing uses are already
bools, but future uses might not be.
2017-07-06 08:53:01 +10:00
Nigel Tao 7bbd39af69 Have writeCtorSignature return an error.
It's redundant, but consistent with all of the other writeXxx methods.
2017-07-02 16:14:31 +10:00
Nigel Tao 187bde7540 Pull out duplicate short_read_src code.
name                          old speed      new speed       delta
lzw_puffs_decode_20k/clang     127MB/s ± 0%    127MB/s ± 0%  +0.40%  (p=0.008 n=5+5)
lzw_puffs_decode_100k/clang    176MB/s ± 0%    176MB/s ± 0%  +0.23%  (p=0.008 n=5+5)
gif_puffs_decode_1k/clang      380MB/s ± 1%    389MB/s ± 2%  +2.26%  (p=0.008 n=5+5)
gif_puffs_decode_10k/clang     136MB/s ± 0%    137MB/s ± 0%  +0.59%  (p=0.016 n=4+5)
gif_puffs_decode_100k/clang    121MB/s ± 0%    121MB/s ± 0%  +0.33%  (p=0.008 n=5+5)
gif_puffs_decode_1000k/clang   124MB/s ± 0%    124MB/s ± 0%  +0.25%  (p=0.008 n=5+5)
lzw_puffs_decode_20k/gcc       150MB/s ± 1%    151MB/s ± 0%  +0.97%  (p=0.008 n=5+5)
lzw_puffs_decode_100k/gcc      207MB/s ± 0%    197MB/s ± 0%  -4.84%  (p=0.008 n=5+5)
gif_puffs_decode_1k/gcc        413MB/s ± 1%    406MB/s ± 1%  -1.71%  (p=0.008 n=5+5)
gif_puffs_decode_10k/gcc       147MB/s ± 0%    158MB/s ± 0%  +7.15%  (p=0.008 n=5+5)
gif_puffs_decode_100k/gcc      136MB/s ± 0%    138MB/s ± 0%  +1.68%  (p=0.008 n=5+5)
gif_puffs_decode_1000k/gcc     139MB/s ± 0%    142MB/s ± 0%  +2.07%  (p=0.008 n=5+5)
2017-07-02 16:11:51 +10:00
Nigel Tao d974988eab Reset coro_state to zero at function end.
This allows us to call that coroutine function multiple times.
2017-07-02 13:59:42 +10:00
Nigel Tao a65e2bacf0 Add mimic tests. 2017-07-02 11:19:23 +10:00
Nigel Tao a58bfc67f1 Add "bench" to puffs' usage message. 2017-07-01 11:26:55 +10:00
Nigel Tao 94dcc14b44 Have writer1 methods honor the limit.
The benchmark numbers are worse, but we need to honor the limit, which
we weren't doing before.

name                          old speed     new speed     delta
bench_lzw_decode_20k/clang    128MB/s ± 0%  116MB/s ± 0%   -9.87%  (p=0.016 n=5+4)
bench_lzw_decode_100k/clang   166MB/s ± 0%  154MB/s ± 0%   -7.18%  (p=0.008 n=5+5)
bench_gif_decode_1k/clang     351MB/s ± 0%  285MB/s ± 1%  -18.73%  (p=0.008 n=5+5)
bench_gif_decode_10k/clang    131MB/s ± 0%  122MB/s ± 0%   -6.70%  (p=0.008 n=5+5)
bench_gif_decode_100k/clang   121MB/s ± 0%  110MB/s ± 0%   -9.37%  (p=0.008 n=5+5)
bench_gif_decode_1000k/clang  124MB/s ± 0%  112MB/s ± 0%   -9.13%  (p=0.016 n=5+4)
bench_lzw_decode_20k/gcc      143MB/s ± 0%  138MB/s ± 2%   -2.94%  (p=0.008 n=5+5)
bench_lzw_decode_100k/gcc     195MB/s ± 0%  175MB/s ± 2%  -10.23%  (p=0.008 n=5+5)
bench_gif_decode_1k/gcc       394MB/s ± 0%  389MB/s ± 3%   -1.42%  (p=0.008 n=5+5)
bench_gif_decode_10k/gcc      137MB/s ± 0%  141MB/s ± 1%   +2.87%  (p=0.008 n=5+5)
bench_gif_decode_100k/gcc     127MB/s ± 0%  128MB/s ± 2%     ~     (p=0.730 n=4+5)
bench_gif_decode_1000k/gcc    130MB/s ± 0%  130MB/s ± 2%     ~     (p=0.730 n=4+5)
2017-06-30 15:30:51 +10:00
Nigel Tao 8be3e9e91f Optimize reader1 calls.
name                          old speed     new speed     delta
bench_lzw_decode_20k/clang    110MB/s ± 2%  128MB/s ± 0%  +16.73%  (p=0.008 n=5+5)
bench_lzw_decode_100k/clang   149MB/s ± 0%  166MB/s ± 0%  +11.40%  (p=0.008 n=5+5)
bench_gif_decode_1k/clang     281MB/s ± 0%  351MB/s ± 0%  +24.55%  (p=0.008 n=5+5)
bench_gif_decode_10k/clang    106MB/s ± 0%  131MB/s ± 0%  +23.71%  (p=0.008 n=5+5)
bench_gif_decode_100k/clang   101MB/s ± 0%  121MB/s ± 0%  +19.87%  (p=0.008 n=5+5)
bench_gif_decode_1000k/clang  103MB/s ± 0%  124MB/s ± 0%  +19.61%  (p=0.008 n=5+5)
bench_lzw_decode_20k/gcc      136MB/s ± 0%  143MB/s ± 0%   +4.61%  (p=0.008 n=5+5)
bench_lzw_decode_100k/gcc     187MB/s ± 0%  195MB/s ± 0%   +4.23%  (p=0.008 n=5+5)
bench_gif_decode_1k/gcc       392MB/s ± 1%  394MB/s ± 0%   +0.67%  (p=0.008 n=5+5)
bench_gif_decode_10k/gcc      129MB/s ± 0%  137MB/s ± 0%   +6.07%  (p=0.008 n=5+5)
bench_gif_decode_100k/gcc     124MB/s ± 0%  127MB/s ± 0%   +2.51%  (p=0.016 n=5+4)
bench_gif_decode_1000k/gcc    126MB/s ± 0%  130MB/s ± 0%   +3.28%  (p=0.016 n=5+4)
2017-06-30 15:04:39 +10:00
Nigel Tao 6fde16afd5 Refactor puffs-gen-c into smaller files. 2017-06-30 10:56:37 +10:00
Nigel Tao 0cba6885fa Spell the "-b" flag fully, as "-bench". 2017-06-30 08:50:12 +10:00
Nigel Tao 0203d2f148 Compile with gcc without -std=c99. 2017-06-29 16:27:04 +10:00
Nigel Tao 480c509214 Add benchmarks. 2017-06-29 14:13:23 +10:00
Nigel Tao 8fe438d0b4 Temporary code to continue this.lzw.decode calls.
This is a hard-coded, short term fix so that GIF decoding works
end-to-end. The long term fix is to have some principled way of
responding to non-fatal status codes such as "short read".
2017-06-23 15:46:08 +10:00
Nigel Tao 8ad82f5e1f Generate C code to set coroutine state. 2017-06-23 15:35:18 +10:00
Nigel Tao bacd2e3b61 Add a PUFFS_COROUTINE_STATE macro. 2017-06-23 14:47:34 +10:00
Nigel Tao 24c8509242 Generate C "goto suspend" statements. 2017-06-23 14:05:30 +10:00
Nigel Tao 9ec2ad7e0c Generate C coroutine switch cases. 2017-06-23 13:55:20 +10:00
Nigel Tao 52b906a261 Generate C code to save and load coroutine state. 2017-06-23 13:03:44 +10:00
Nigel Tao 19960c40f8 Generate a C coroutine switch.
The technique is similar to
https://www.chiark.greenend.org.uk/~sgtatham/coroutines.html

Generating "case NNN" targets is still a TODO.
2017-06-23 11:42:35 +10:00
Nigel Tao e21c68a006 Generate C struct fields to save coroutine state. 2017-06-23 10:49:56 +10:00
Nigel Tao 09c59dfade Tweak puffs_base_limit1 field and type names. 2017-06-23 10:14:44 +10:00
Nigel Tao 0088c7e257 Have read_u8 honor the limit. 2017-06-23 10:01:31 +10:00
Nigel Tao 7a3ddbc025 Add a "limit" keyword. 2017-06-22 17:32:56 +10:00
Nigel Tao 1d39b1e173 Add a dst argument to std/gif's decoder.decode_id. 2017-06-22 15:41:10 +10:00
Nigel Tao 497a8776c2 Generate C code for "this.lzw.set_literal_width". 2017-06-22 14:56:29 +10:00
Nigel Tao 1a42706bf1 Add a reader1.skip32 method. 2017-06-22 13:09:53 +10:00
Nigel Tao 8134c89786 Decode the rest of a GIF.
Passing an Image Descriptor's block data payload to the LZW decoder is
still a TODO.
2017-06-22 11:25:07 +10:00
Nigel Tao 17e47fbc1e Add reader1 and writer1 base types. 2017-06-18 11:50:34 +10:00
Nigel Tao dc71a93dcf Re-introduce type/bounds checking to "puffs gen".
This rolls back some of:
f23a808 Remove some redundant cmd/puffs/gen.go code.
3258c1e Remove check.FlagsOnlyTypeCheck.

The language-specific code generators, such as puffs-gen-c, have to do
type and bounds checking anyway, to generate the right types and
function calls. Accordingly, I thought that the type and bounds checking
in the language-agnostic cmd/puffs/gen.go was redundant, and removed it.

However, puffs-gen-c just works with stdin and stdout, and the comments
have been stripped from its input, so it cannot print either the right
filename or right line number for any errors.

We therefore need to do type and bounds checking in cmd/puffs/gen.go
too, in order to produce actionable errors.
2017-06-18 10:17:09 +10:00
Nigel Tao 8d1bf15d70 Add a decoder.decode_lsd skeleton. 2017-06-18 09:45:09 +10:00
Nigel Tao 2543999c61 Remove a comment typo. 2017-06-18 09:29:45 +10:00
Nigel Tao f23a80820c Remove some redundant cmd/puffs/gen.go code. 2017-06-18 09:26:30 +10:00
Nigel Tao 7c85ed04f9 Check that __WORDSIZE <= 64. 2017-06-18 09:21:58 +10:00
Nigel Tao 6904068251 Remove check.FlagsOnlyTypeCheck.
The puffs-gen-foo commands might need to bounds check anyway, to e.g.
generate different code paths depending on whether a u64 typed
expression fits in a usize's range.
2017-06-18 09:17:58 +10:00