eden-emu/opus
15
0
Fork 0
mirror of https://github.com/xiph/opus.git synced 2025-06-03 00:57:43 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
456 commits 284 branches 59 tags 26 MiB
Commit graph

456 commits

This branch
This branch
All branches
Author SHA1 Message Date
Jean-Marc Valin
8bdbbfa18d Support for sparse GRU B input matrices 
Only on the C side, no sparse GRU B training yet
 2021-07-16 03:07:26 -04:00
Jean-Marc Valin
4c0e224865 Model update 
Weights are the same, but they are dumped to C differently.
 2021-07-15 16:12:42 -04:00
Jean-Marc Valin
c74330e850 Pre-compute GRU B conditioning 
Adapted from PR: https://github.com/mozilla/LPCNet/pull/134
by zhuxiaoxu <zhuxiaoxu@ainirobot.com>
but had to be reworked due to previous weight quantization changes.
 2021-07-15 16:06:56 -04:00
Jean-Marc Valin
0d53fad50d Using np.memmap() to load the training data 
Makes loading faster
 2021-07-14 13:47:23 -04:00
Jean-Marc Valin
5a51e2eed1 Adding command-line options to training script 2021-07-13 03:09:04 -04:00
Jean-Marc Valin
1edf5d7986 README.md update 2021-07-11 03:46:25 -04:00
Jean-Marc Valin
4298f2f9e1 Adding support for SSE2 and SSSE3 2021-07-11 03:36:20 -04:00
Jean-Marc Valin
116bcb38fb Adding SSE 4.1 for older platforms 
AVX without AVX2 should now work again too.
 2021-07-10 14:08:01 -04:00
Jean-Marc Valin
3e223e6015 Fixes Python inference for the binary probability tree 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
f8f12e7f3c NEON float->char conversion (same as the AVX2 version) 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
a1079c2ce3 Again, same conversion as 3206cec, for NEON 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
7d8b00f11d Sampling directly from the logit 
Avoids having to compute a sigmoid
 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
e8f70128d5 same conversion cleanup as 3206cec for sgemv_accum8x4() 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
7cef98ec8c Minor optimization: merging all 3 embeddings 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
714380e71b More manual unrolling 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
006556036a Cleaning up the sparse GRU 
It no longer overwrites its input vector
 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
44fe055682 cleanup float<->int conversions 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
60d6eab63d Doing a bit of unrolling to speed things up 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
3e7ab9ff87 update model 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
54abdb6f5d Sparse matrix indexing optimization 
The 4* is now stored in the table to avoid computing it in the loop
 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
2681822c18 update model 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
d332100808 Representing output pdf as binary probability tree 
Saves on the MDense/softmax computation since we only need to compute
8 values instead of 256.
 2021-07-10 01:59:49 -04:00
Jean-Marc Valin
c151fc1853 Merge branch 'exp_improved_simd2' 2021-06-30 18:56:04 -04:00
Jean-Marc Valin
f0ce43389a Update test_lpcnet.py, remove old TF1 code 2021-06-30 18:54:27 -04:00
Jean-Marc Valin
d428b0d32a Update model 2021-06-30 18:27:31 -04:00
Jean-Marc Valin
8c4b88cfab Using a bisection search for sampling 2021-06-30 18:14:12 -04:00
Jean-Marc Valin
e35441f2cc Faster activation functions for AVX 
Using rational function approximation for tanh() and sigmoid.
 2021-06-29 04:05:48 -04:00
Jean-Marc Valin
5571ef1b8e minor optimization: removing some copying 2021-06-26 01:27:03 -04:00
Jean-Marc Valin
d61f7e00f8 Fix missing transpose in the sparity code 
CuDNNGRU and GRU don't use the same weight format
 2021-06-25 13:43:37 -04:00
Jean-Marc Valin
ca0a43bee9 Update README.md 2021-06-24 17:47:51 -04:00
Jean-Marc Valin
c1535c8ccf Adding option to disable int8 dot products 2021-06-24 17:31:05 -04:00
Jean-Marc Valin
0b9f6bab81 Remove unnecessary mask in exp() approximation 
This isn't necessary since valid exponents can't flip the sign bit
 2021-06-21 01:34:38 -04:00
Jean-Marc Valin
ae2ae5ead6 Remove useless multiply by one 
See bffdcee95 (commitcomment-46372726)
 2021-06-21 01:30:51 -04:00
Jean-Marc Valin
c7ba313a67 Adding extra constraint to avoid saturation for SSE/AVX2 
When implementing using SSSE3 or AVX2, our dot products can saturate
if two adjacent weights sum to more than 127.
 2021-06-18 17:39:35 -04:00
Jean-Marc Valin
237245f815 Support for multi-GPU training 
Not sure why CuDNNGRU doesn't get used by default, but we need
to explicitly use it to get things to run fast.
 2021-06-18 13:20:43 -04:00
Jean-Marc Valin
ebc9483b4c update model 2021-02-01 01:07:35 -05:00
Jean-Marc Valin
79980b2044 Minor update to training scripts 2021-01-18 02:13:52 -05:00
Jean-Marc Valin
20fea538c2 more reasonable noise 
was increased too much in 713d53e8a
 2021-01-17 21:39:42 -05:00
Jean-Marc Valin
b9c230b346 Add NEON intrinsics 2021-01-16 02:11:22 -05:00
Jean-Marc Valin
b214e684c1 Neon WIP: Compiles but very slow 2021-01-16 02:11:21 -05:00
Jean-Marc Valin
8c3fe6f31d Cleaning up float version 2021-01-16 02:11:21 -05:00
Jean-Marc Valin
40b9fd0a75 Fix some quantization issues 2021-01-16 02:11:21 -05:00
Jean-Marc Valin
83657d0e43 Dot product AVX2 code for non-sparse multiply 2021-01-16 02:11:21 -05:00
Jean-Marc Valin
1707b960de cleanup, add signed-unsigned biases 2021-01-16 02:11:21 -05:00
Jean-Marc Valin
40b309d92b WIP: 8-bit SIMD for GRU B 2021-01-16 02:11:21 -05:00
Jean-Marc Valin
e695355ba5 some cleanup 2021-01-16 02:11:20 -05:00
Jean-Marc Valin
06489b42dd oops, fix number of columns 2021-01-16 02:11:20 -05:00
Jean-Marc Valin
d87f974431 Vectorizing conversion 2021-01-16 02:11:20 -05:00
Jean-Marc Valin
6b582edbed WIP: remove scalar code from AVX2 code 2021-01-16 02:11:20 -05:00
Jean-Marc Valin
be392e3857 WIP: Got some AVX2 code working 2021-01-16 02:11:20 -05:00