BRL-CAD/mkl-dnn
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
6,976 Commits 3 Branches 36 Tags
cc77eef809d0331b245eb21a7956d507505700aa
Commit Graph
38 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Andrzejewski, Kamil cc77eef809 benchdnn: reorder: add new configuration to compensation tests 2021-06-09 17:03:08 +07:00
Andrzejewski, Kamil 5d3d1e1874 cpu: reorder: enable unsupported compensation reorder when src_tag=bxa 2021-06-09 17:03:08 +07:00
Andrzejewski, Kamil 4f05b76bb7 cpu: reorder: swap duplicate configuration to the corrected one 2021-06-09 17:03:08 +07:00
Kamil Andrzejewski f7b90c7b22 git: update gitignore 2021-02-09 10:01:01 +01:00
Andrzejewski, Kamil daff29808b api: eltwise: add hardswish algorithm 2021-01-24 23:19:29 -08:00
Andrzejewski, Kamil f808316405 gpu: add mish eltwise algorithm 2021-01-18 08:31:45 +01:00
Andrzejewski, Kamil 70d8c146a6 cpu: x64: eltwise: add mish jit implementation 2021-01-18 08:31:28 +01:00
Andrzejewski, Kamil 6a566a56d1 doc: add mish eltwise algorithm 2021-01-18 08:28:17 +01:00
Andrzejewski, Kamil 3c4ca3fda0 benchdnn: add mish eltwise algorithm 2021-01-18 08:28:17 +01:00
Andrzejewski, Kamil 7273c14f15 common: add mish eltwise algorithm 2021-01-18 08:28:17 +01:00
Andrzejewski, Kamil 999d511f13 cpu: bnorm: add support for nspc format in jit tbb kernel 2020-12-16 03:07:53 -08:00
Andrzejewski, Kamil 800a200aeb cpu: bnorm: refactor to make code cleaner and more readable 2020-12-16 03:07:53 -08:00
Andrzejewski, Kamil c16a19ebfb cpu: eltwise: fix for issue when exp(x) is near to max value 
x*log2e can exceed 128 when exp is near to max value for
polynomial with arg reduction from [-log(2)/2; log(2)/2].
This number is used in next computations to count 2^n,
where n can be 128 as it was presented earlier. 2^128 is not
representable in fp32, to around this problem 2^128 can be
divided into two numbers 2 and 2^127 which are
representable in fp32. This change fix issues in all algs
which have in their formula exp(x) i.e. gelu_erf, elu,
soft_relu, logistic, exp.
 2020-11-16 00:26:05 -08:00
Andrzejewski, Kamil 7cc156de5b revert: cpu: eltwise: fix for issue when exp(x) is near to max value 2020-11-16 00:26:05 -08:00
Andrzejewski, Kamil e5ebb67d75 gpu: add clip_v2 eltwise algorithm 2020-11-12 00:29:08 -08:00
Andrzejewski, Kamil 8242f55aa9 cpu: x64: eltwise: add clip_v2 jit implementation 2020-11-12 00:29:08 -08:00
Andrzejewski, Kamil 24ea57a156 doc: add clip_v2 eltwise algorithm 2020-11-12 00:29:08 -08:00
Andrzejewski, Kamil ac0bb521e8 benchdnn: add clip_v2 eltwise algorithm 2020-11-12 00:29:08 -08:00
Andrzejewski, Kamil 997ff52af1 common: add clip_v2 eltwise algorithm 2020-11-12 00:29:08 -08:00
Andrzejewski, Kamil 10b1b8cd95 cpu: eltwise: fix for issue when exp(x) is near to max value 
Change the arg reduction from [-log (2) / 2; log (2) / 2]
to [0; log (2)] to be able to compute "trunc (x * log2ef)"
and not exceed 128. This change fix issues in all algs
which have in their formula exp(x) i.e. gelu_erf, elu,
soft_relu, logistic, exp.
 2020-11-02 03:53:01 -08:00
Andrzejewski, Kamil 8bc2413f8e cpu: resampling: use std::function instead of pointer to function 2020-10-16 07:40:30 +02:00
Andrzejewski, Kamil 9ca20657ad cpu: resampling: change data distrubution between threads to speed up ncsp format 2020-10-16 07:40:26 +02:00
Andrzejewski, Kamil e68682cfd9 cpu: resampling: remove old kernel for avx512 for fwd path 2020-10-16 07:40:23 +02:00
Andrzejewski, Kamil afa2507d6e cpu: resampling: introduce new jit uni kernel for fwd path 2020-10-16 07:40:17 +02:00
Andrzejewski, Kamil 513ab934d5 cpu: pooling: unification memory tag kind 2020-10-14 13:54:57 +02:00
Andrzejewski, Kamil 14c9de4094 benchdnn: binary: add skip cases for different data types on cpu 2020-10-12 05:45:43 -07:00
Andrzejewski, Kamil c492165449 benchdnn: inputs: binary: add additional test cases for post-ops feature 2020-10-12 05:45:43 -07:00
Andrzejewski, Kamil 1b6078d801 cpu, tests: resampling: improve the accuracy 2020-08-07 00:37:09 -07:00
Andrzejewski, Kamil 9172a741a7 cpu: bnorm: fix for illegal division by 0 2020-07-23 00:15:57 -07:00
Andrzejewski, Kamil b0fd13ac7c cpu: bnorm: turning on the barrier when SP*N == 1 in ncsp format 2020-07-22 23:52:51 -07:00
Andrzejewski, Kamil 5aad455b2c cpu: reorder: casting input val to fp32 before substract zero point 2020-07-17 02:27:05 -07:00
Andrzejewski, Kamil 00f2d5ff38 cpu: pooling: refactor 2020-07-03 14:56:21 +02:00
Andrzejewski, Kamil 1f205df57a cpu: pooling: support tail processing by channels in nhwc jit pooling 2020-07-03 11:51:50 +02:00
Andrzejewski, Kamil adb616931e benchdnn: bnorm: fix for division by 0 during cmp when N * SP == 1 2020-06-16 07:46:13 -07:00
Andrzejewski, Kamil 12b2bc4330 cpu: resampling: refactor simple impl (moving defs from hpp to cpp, etc) 2020-05-14 23:18:12 -07:00
Andrzejewski, Kamil 5a235144d4 cpu: resampling: add bfloat16 support to simple impl 2020-05-14 23:18:12 -07:00
Andrzejewski, Kamil fe95d9cd84 cpu: resampling: add jit_avx512_common impl for fwd and bwd prop 2020-05-14 23:18:12 -07:00
kandrzej 176458282f cpu: resampling: added missing memory tags for simple_resampling 2020-04-27 12:58:04 +02:00