[GNA] NCHW to NHWC layout transformations (#18120)

* Squashed commit of the following: GNAConv+ MaxPool commit 51e7254eab Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Apr 18 18:09:46 2023 +0200 reset json schema commit 6d87c40703 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Apr 18 18:02:53 2023 +0200 Exclude wrong changes commit 30dd9aef86 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Apr 18 17:56:54 2023 +0200 clang format commit dad7151b37 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Apr 18 17:47:17 2023 +0200 Init commit * Squashed commit of the following: GatherSinking binary elementwise commit 960f7ebaed Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 13:36:35 2023 +0100 fix year commit 860ae33f38 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 13:29:05 2023 +0100 add docs commit bd35e57c60 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:58:55 2023 +0100 cleanup commit 3fed498f03 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:52:49 2023 +0100 clang fixes + remove unneeded functions commit 4417a13bad Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:50:03 2023 +0100 fix unit tests execution commit 48f20927af Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:33:03 2023 +0100 initial * Squashed commit of the following: common debug merge commit 9cdddb337972c2ebecb3ed81f2613b96ae034538 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri May 12 14:05:34 2023 +0200 clnag format commit f093baa7708bd9e9527277337e52104b4e82744d Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 19 15:27:58 2023 +0200 fix ts_concat ts_split commit 1c8baba5bb4c1533f4470505fe5a0dce6e4bc45e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 19 13:26:30 2023 +0200 fix ts_split commit 043af9aad51b218c8e27150516e3824cea682ca3 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 18 19:39:15 2023 +0200 remove debug commit 7c03af67f018d5eea2c39d01bf2d68b815cde390 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 18 19:37:34 2023 +0200 fix ts_concat commit dc2f00a1e77427b0ae243f5ef7606eac4cdf8051 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 18 13:15:50 2023 +0200 implemented ts_split through gather commit 311045f78d357b4f1129a4545da9092435313664 Merge: 06fee24288 7ca64ec9f7 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Apr 17 19:19:05 2023 +0200 Merge branch 'gna_layout_debug_fix' into gna_layout_debug commit 7ca64ec9f704c2228808212e267fb387c08e4921 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Apr 17 19:06:44 2023 +0200 fix commit 81c6217c68fe6f945ab83ae8b6ed64d23415a62e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Apr 17 18:29:32 2023 +0200 fix commit 06fee24288be846209e87fbd40aab7cad2775559 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Apr 17 16:03:34 2023 +0200 fix RemoveOutputsProcessing name to emplace commit 83344b2723b4db919862e7519369593710d06de9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Apr 17 10:35:59 2023 +0200 fix commit cc7d6bdea26df1d2e3195bf82e62193442723995 Merge: d916fa40f9 83d51b5fe4 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 13 13:32:44 2023 +0200 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit d916fa40f98d5d253e445929319f06b638d2f430 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 13 13:32:02 2023 +0200 remove debug commit f8b1fac6f94577059f729ac7c31594b4f93ba4ae Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 13 13:31:12 2023 +0200 fix + remove debug commit 710d004eda3063374b99c3a78a37f7af2b84d1e8 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 13 11:54:36 2023 +0200 fix commit 83d51b5fe485d96c76bc77ace671491c5bfb1223 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Apr 12 19:06:15 2023 +0200 Pattern Reshape-Transpose commit 3e33e51edff7762112f00d29281a59a6cb16ea83 Merge: a5042ebce7 74c68cbf4f Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Apr 12 16:56:24 2023 +0200 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit a5042ebce79b0fc6a5db749166a4784aec9f0d2a Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Apr 12 16:56:00 2023 +0200 Reverted check of concat with interleaved inputs commit 74c68cbf4fd1206d5e67326a4cf0dc290fbeb574 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 12 16:53:17 2023 +0200 use ts_fuse instead of general after all commit 49ac734247ef6f73ce2ca096f1bd03c027f3c338 Merge: 79d23d893f 1dbea2cb3c Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Apr 12 15:02:35 2023 +0200 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 79d23d893fd585f253eec8d3e10198e5b75ed2da Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Apr 12 15:02:08 2023 +0200 Fixed ConvertMatmulToPointwiseConvWithFqNeg commit 1dbea2cb3c948d2703a2e97957b25ebd2bd30db6 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 12 14:29:22 2023 +0200 fix pipeline commit d73effb2c1905754a9f8636e0a762885224ca44f Merge: e2b5440ad1 d5b7a7c984 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 12 14:01:28 2023 +0200 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit e2b5440ad11eea5aa456146fa12a5f07223c2f43 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 12 14:01:25 2023 +0200 fix ts_split bug when we some split outputs are not connected commit d5b7a7c98477df74fd6e1a8481c8bc513854001b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Apr 12 13:57:25 2023 +0200 Corrected limit conditions commit 7de41f2072420f3d1cb1f933991853ba787a8f42 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 12 13:04:42 2023 +0200 remove debug commit ad2584aa23265c3cdb6370d1b87a8c0af89c17ad Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 11 18:48:49 2023 +0200 use TSConcatForward TSSplitBackward commit 7c9fabb239415e3f1af01058aaca8b9bc63d6c26 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 11 18:33:03 2023 +0200 fix ts_split commit da1488d21ccb649e75d0e61d424bf8e8f5d9e8f2 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 11 14:19:26 2023 +0200 fixed ts_concat commit 2014b727ed387ca2035de5ec144f1beb67c32dd9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 6 11:42:17 2023 +0200 add ts_split commit 7316a17131a6cb0a8c25061a6b12afdee004661d Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 6 11:42:00 2023 +0200 remove debug commit 05c9b9524781f2017d6c5f88b793c210dfd2ab16 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Apr 6 11:41:10 2023 +0200 remove debug, turn off new transformations commit 8962424b907f7fec2ddfe32912f8da77c13461e4 Merge: f98adb6dc4 99090c2090 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 5 16:50:41 2023 +0200 Merge branch 'gna_layout_debug' into gna_layout_debug_concat_split commit f98adb6dc40b4e368e97b58cc28ad3554fb54923 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 5 16:50:17 2023 +0200 remove debug commit 9dddf3ada89b65da3191f647bf2f17af6d8a092a Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 5 16:44:34 2023 +0200 fix ts_concat sigsegv commit 190010f4dfca82bb4b5a3e045e6015b8d1db8491 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 5 13:15:53 2023 +0200 fix problem with transpose backward sinking after gna::ts_concat commit 99090c2090e3a121241a54a5b8da290ba5272b11 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 5 12:52:13 2023 +0200 fix bug GatherSinkingTransposeReshapeForward map:at exception commit a6f65aae2d94a63f60440d9fcdeeb31af548b055 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Apr 5 12:51:19 2023 +0200 add ts_concat commit d9497eadad5c58ec9218692a08da54c3d7fd3d40 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 13:40:51 2023 +0200 add code from Misha commit a5fe706b2bf92bcc1f3f8c435663551a0770188c Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Apr 4 13:00:47 2023 +0200 TS concat callback draft commit 8f244188c4d0ceef11850e118fd0ad23fcdac0b8 Merge: 5591091e05 a5fe706b2b Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 12:57:23 2023 +0200 Merge branch 'gna_layout_debug' into gna_layout_debug_concat_split commit 5591091e05a32b8e3566ec70490cbd373f827910 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 12:56:48 2023 +0200 fix commit 7f9eb3c02c57cbef5b32efa5d61cf7082ca6b621 Merge: a98a353f00 9190b8ccff Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 11:20:26 2023 +0200 merge commit a98a353f00cdb1650edfe0f5f24a10111c0319b7 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 11:19:43 2023 +0200 fix commit a8ca19419f7b1e16064c8cb53ed836370da8af5e Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Apr 4 11:14:44 2023 +0200 Fixed weights reducer commit 9190b8ccffd1a125cb97d0e339f0820f1c565d3c Merge: 279f76287e 4ce59b6885 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 10:46:07 2023 +0200 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 279f76287eb7e656c3dd743a228839c16470057a Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 10:46:03 2023 +0200 use transformation_callback in ts_concat commit 3fabc31ed9462f86ae173aba927caf9c9c3fb539 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Apr 4 10:44:42 2023 +0200 fix commit d942074be54a99147c893eb78e878cc4428be5ad Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Apr 3 18:29:02 2023 +0200 debug commit 4ce59b68851f71a049b34d16008a6ed82cc91d2b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Apr 3 17:49:45 2023 +0200 Skipped supported transposes in Gather transform commit 3d57040037493aa61ffb1d159e76567ca493dd33 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Apr 3 12:16:58 2023 +0200 Extended FQ detection commit bcc317e80d1f9ed6d6f6aa5c77ff5e58e31e84db Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Apr 3 10:11:48 2023 +0200 Gather sinking with FQ Backward commit b1dc3c44f3cca1035ce4bd3ff05f1f78a7add1ce Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 31 10:48:38 2023 +0200 Fixed typo commit 76b2e162156e20a1ade9afc893e9874b8c66284d Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Mar 28 15:08:24 2023 +0200 FQ fix commit 99df8e6164cad78a25e34ea752da9264c7536188 Merge: 07cdff9a44 8894dfc88b Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 27 18:03:54 2023 +0200 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 07cdff9a44585026bbd58fc1c33b1e67d47bd225 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 27 18:03:51 2023 +0200 add GatherSinkingMatmulBackward commit 8894dfc88b22736b1157cd75e6b48ad14a4a1137 Merge: 0e0dcc6a53 67a4035902 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 24 14:28:39 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 67a4035902 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Mar 24 13:59:20 2023 +0100 remove debug commit 04a715267a Merge: 7b95c90df8 1163b926ee Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Mar 24 13:56:05 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 7b95c90df8 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Mar 24 13:54:49 2023 +0100 fix gather_sinking_matmul commit 0e0dcc6a53ce660b10c27d7e1d48de260e87e127 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 24 10:43:07 2023 +0100 Disabled ReshapeTransposeSubstitute commit daaeeecfac3ca7b1763279e62455d007245bcdf8 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 24 10:41:56 2023 +0100 Fixed squeeze helper commit 1163b926ee Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Mar 23 16:09:59 2023 +0100 Corrected Concat checks commit 06d6fbf0e8 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 23 15:28:00 2023 +0100 initial commit 1ca78f643e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 23 15:27:21 2023 +0100 debug commit 3c5aa8c53d Merge: 01de25e3c6 aaba7740d4 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Mar 23 15:13:02 2023 +0100 Merge branch 'itikhono/ts/slice' into gna_layout_debug commit 01de25e3c6 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Mar 23 13:35:35 2023 +0100 detect matmuls commit 45c1f800f3 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Mar 23 12:34:28 2023 +0100 Removed do_rotate flag processing commit aaba7740d4 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 23 10:01:58 2023 +0000 codestyle commit 979d5cc669 Merge: 67c1b9daad a66868b463 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 23 17:48:05 2023 +0400 Merge branch 'itikhono/ts/slice' of https://github.com/itikhono/openvino into itikhono/ts/slice commit 67c1b9daad Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 23 17:47:50 2023 +0400 resolve review comments commit 8b81867d18 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Mar 23 10:15:42 2023 +0100 Reverted GetConvData helpers commit d78f7e7fc6 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 22 17:29:02 2023 +0100 add gather sinking split commit e1ff589d3e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 22 17:07:12 2023 +0100 ad Reshape to supported gather sinking nodes commit 894defdcc9 Merge: 31d7af368e 4fd047deec Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 22 16:00:51 2023 +0100 Merge branch 'gna_layout_debug' into gather_sinking_reshape commit 31d7af368e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 22 16:00:00 2023 +0100 add gather_sinking_reshape commit 4fd047deec Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 22 15:29:59 2023 +0100 Fixed squeeze condition commit d105cfcc68 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 22 14:55:10 2023 +0100 Fixed Gather indexes commit a66868b463 Merge: bf9bc8628b b70e56d110 Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Wed Mar 22 01:53:10 2023 +0400 Merge branch 'master' into itikhono/ts/slice commit bf9bc8628b Merge: 31bd608ac4 3d508c7b36 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 21:41:57 2023 +0000 Merge branch 'itikhono/ts/slice' of https://github.com/itikhono/openvino into itikhono/ts/slice commit 31bd608ac4 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 21:41:37 2023 +0000 fix TSSqueeze transformation commit 02abf9b1f0 Merge: 4481041f86 5c249d98d9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 21 19:48:06 2023 +0100 Merge branch 'transpose_sinking_fakequantize' into gna_layout_debug commit 5c249d98d9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 21 19:47:27 2023 +0100 fix commit cdfd77a415 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 21 19:43:43 2023 +0100 rebase fixes commit e4207c4d6b Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 20:21:42 2023 +0400 fix unit tests, revert changes for TSSlice transformation commit 4481041f86 Merge: eac5d4beb4 b1099c1c40 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 21 16:16:56 2023 +0100 Merge branch 'transpose_sinking_fakequantize' into gna_layout_debug commit b1099c1c40 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 2 14:38:43 2023 +0100 initial commit 0f17c5f714 Merge: 981e8ad3c0 db220f11ee Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 18:24:08 2023 +0400 Merge branch 'itikhono/ts/slice' of https://github.com/itikhono/openvino into itikhono/ts/slice commit 3d508c7b36 Merge: 9720a28fa6 0893efe073 Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Tue Mar 21 17:07:42 2023 +0400 Merge branch 'master' into itikhono/ts/slice commit 9720a28fa6 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 12:35:23 2023 +0000 codestyle commit eac5d4beb4 Merge: faf2f253c6 db220f11ee Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Mar 21 12:45:59 2023 +0100 Merge branch 'itikhono/ts/slice' into gna_layout_debug # Conflicts: # src/common/transformations/src/transformations/transpose_sinking/ts_slice.cpp # src/common/transformations/src/transformations/transpose_sinking/ts_squeeze.cpp # src/common/transformations/src/transformations/transpose_sinking/ts_unsqueeze.cpp commit faf2f253c6 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Mar 21 12:12:18 2023 +0100 Squeeze shapes when detect gather commit 7b1bbf77ee Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 20 19:48:56 2023 +0100 fix /gather_sinking_transpose_reshape commit e6650cfd1c Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Mar 21 12:06:01 2023 +0100 Corrected Convolution checks commit 83f4428f48 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 20 18:01:35 2023 +0100 Corrected reshape paterns commit 0552dfe537 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Sun Mar 19 18:39:01 2023 +0100 Merged the latest transpose sinking changes commit 56a5e74bf4 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Sun Mar 19 18:21:06 2023 +0100 Revert "Squashed commit of the following:" This reverts commit 227274e70696def4a718e65799edb3b50c643c65. commit fd8effeef5 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Sun Mar 19 18:17:25 2023 +0100 Squashed commit of the following: commit f1b49907a34c94ca91fad643e8949d2ae38b2f44 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 17 15:44:24 2023 +0100 Removed unnecessary transposes commit e57f26cf775d74a61d6e505f785211bdf2577405 Author: Ilya Lavrenov <ilya.lavrenov@intel.com> Date: Thu Mar 9 12:32:02 2023 +0400 Fixed PT FE compilation with clang on macOS (#16173) commit 3461b52d79f7895df554e7ee023ae7a6448e7b42 Author: Mateusz Bencer <mateusz.bencer@intel.com> Date: Thu Mar 9 07:22:06 2023 +0100 [ONNX FE] Improved a method of operators registration (#15990) * initial version of implementation * styles applied * fixed and registration * add more unit tests * fixed and in legacy opset * review remarks * refactor of version name range commit 1044df5afa63a052750647f57202ce8b7da1906b Author: Jade Cho <jade.cho@intel.com> Date: Thu Mar 9 14:35:31 2023 +0900 [GPU] Fix stable diffusion failure (#16052) * [dGPU] Enable stable diffusion + Prevent to fuse swish into oneDNN reorder. + Makes concat explicitly if batch size is greater than 1 and the siblings are oneDNN impl. commit da1c5cb763b301b050f3f52f362e172c1a42758b Author: Andrew Kwangwoong Park <andrew.park@intel.com> Date: Thu Mar 9 08:36:43 2023 +0900 [GPU] Added shape agnostic Pad kernel implementation (#16160) Signed-off-by: Andrew Park <andrew.park@intel.com> commit 03b96e1d3b41ddca01de7f4117836e5708bd0b7f Merge: ca9e0f6d41 c3ecce0a7e Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 17 11:18:09 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit ca9e0f6d41b95e1653d07c03684efb170d856429 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Fri Mar 17 11:14:36 2023 +0100 Fixed validator of 1D convolution commit c3ecce0a7ea2944b42290c21535b4600a406a1ce Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 16 19:03:10 2023 +0100 fix conditions on mvn and slice commit a6f920ed468943a1c669fdff8b8b0217281ee302 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 16 17:43:35 2023 +0100 fix tests with Slice commit 64b1fa67f5a0ac821d1af5cb132cfd0575041aed Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 16 17:33:38 2023 +0100 fix debug commit ca122e890e1f1590655e9e776d4c858dc3433890 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 15 16:14:49 2023 +0100 move RemovePermutationsNHWCToNCHWPass under is_ngraph_passes_used commit ffa21d987b5f7bc7d374e825ffa6bf4dff7679eb Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 15 13:02:12 2023 +0100 Removed ConvIE code commit 23ca385ba601a06985ea218e53c5a05cdd3b9b87 Merge: 2acd7c4bff 90b37d771b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 15 11:47:57 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 2acd7c4bff4422c95f6b8fb511287110ceb8e99d Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 15 11:45:46 2023 +0100 Disabled sinking for networks without Convolutions commit 90b37d771b9cb9df8d0e1a5d72cb7acc8d01634f Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 14 14:30:37 2023 +0100 fix gather_sinking_transpose_reshape - check if flatten/unflatten the last dimension commit 41e1b72ba9d44a8bd1142735a22feef39e816fcc Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 13 18:26:27 2023 +0100 fix remove output Gather commit 1949dc779a2261686a266bf13fa558078dcce83c Merge: 30db0ee2ce b6fe16e051 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 13 18:13:29 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 30db0ee2ceacf6fb5692421c5647e9ef27794dfb Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 13 18:12:55 2023 +0100 fix: add GatherSinkingTransposeReshape commit b6fe16e051c62cf1b8a1ddf511b2a791587578fc Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 13 17:23:54 2023 +0100 Reverted C++17 code commit fee42e07f7b6861894541cf1d552f82727837a5d Merge: 95330967fc 98ead2d3bc Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 13 16:36:36 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 95330967fc0c39377e33973adeef8c441bb4c905 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 13 16:36:33 2023 +0100 add ReshapeTransposeSubstitute commit 98ead2d3bcc971585d74532ae1263cd7eba8f5b4 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 13 16:33:40 2023 +0100 Fixed FQ connected to Transpose and Inputs commit 6cddb6389e2522c5d9abe262d0a580c01cab4bf8 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 13 16:33:00 2023 +0100 Added add with convolution fusion commit dc705e330f8f23ead78d2e4026eff118028d3dc3 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 10 19:54:19 2023 +0400 TransposeSinking: add support for Slice op commit aba59904c9d95334d6dbea1bdc61971556236d75 Merge: 3dd79763a4 9844f33807 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Mar 10 15:35:00 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 9844f33807c0927b813f90a784a77ee7bd725b57 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Mar 9 14:03:12 2023 +0100 Removed debug commit 3dd79763a4ddc31f7a1838070acf341b610c75ea Merge: cf3154b4f5 8e010d2cba Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 8 18:27:05 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 8e010d2cbab5c338ef877e82641b1e813c88bea0 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 18:26:04 2023 +0100 Corrected condition of unsupported permute commit cf3154b4f5b0d1f793d1b96e97dec6e440b4fd93 Merge: 7c0b042e97 95dcc37315 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 8 18:07:37 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 95dcc3731577db62e42df1c71f2d97068809223e Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 18:07:15 2023 +0100 Removed legacy TransposeWeightsFromNCHWToNHWCPass commit 7c0b042e9763a7382b9ba2706be00681a81f49f6 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 8 17:58:02 2023 +0100 remove debug commit 06730afd016185154fbf6751e00da3673b9ce46b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 17:57:57 2023 +0100 Removed debug code p.2 commit 37660a9721d103144d88d2429ad0f399d20aef32 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 17:51:36 2023 +0100 Removed debug code commit aa8d2abe3a5e169fe9ebb48fe94845bb4aa1b9f1 Merge: 02974d7a12 87194dab68 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 17:50:28 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 02974d7a127caf86a4d46701734cb00fa36439ff Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 17:50:20 2023 +0100 Removed legacy parts commit 87194dab685aa31fe40624381cecfb40f604a6c8 Merge: 225e39d499 235a14301d Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 8 17:31:21 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 225e39d499ed34b7fcd5097c2bfdc86d0d56dc86 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 8 17:31:19 2023 +0100 fix convert_function_to_cnn_network.cpp not using friendly names; fix transpose_nchw update last node name commit 235a14301d1be714e7e0ca074a74ed30120cd327 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Mar 8 17:17:23 2023 +0100 Fixed validation of gna convolution commit dafd9191f52d61b713bb4eec95d63f0a8b0dacfb Merge: cf0e43734b 3a63cb5ad8 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 7 16:27:18 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit cf0e43734ba76a7324a1f0975077b798767f6445 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Mar 7 16:26:45 2023 +0100 fix transpose_nchw friendly names set commit 3a63cb5ad8ac449c909d38743650ef5c514d46c4 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Mar 7 15:41:17 2023 +0100 Extended unsupported transpose condition commit 8b3171c2e60bbdb2ada1765d12e67e8039e17005 Merge: 3385605a9e 69f8479260 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 19:53:01 2023 +0100 Merge branch 'gna_layout_debug' of https://github.com/evkotov/openvino into gna_layout_debug commit 3385605a9e08fc32ab0cbd5de702cb736229d43a Merge: 8d4fa609a6 d3493c48ae Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 19:43:08 2023 +0100 Merge branch 'gna_convolution_maxpool' into gna_layout_debug commit 8d4fa609a67230c89c758dc3aebf22859dcc9498 Merge: 506128f16c 52b9332306 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 19:42:53 2023 +0100 fix merge commit 52b93323062f5d5938af2bf8341a44092380b81b Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 19:41:52 2023 +0100 fix build commit 0e2629313690bb97dde765b10d075cba2cf6278a Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 2 14:59:07 2023 +0100 add binary and gather_fuse commit 0ec73f0239e21d25233cd676818f2e873fdf2f81 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:45:19 2023 +0100 clang commit 28b168c2b4357365bad64983b6f59202726c0a9e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:43:36 2023 +0100 remove debug commit fa2a06fb15b5288746deb3b8947b99cdbba4bba7 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:41:39 2023 +0100 add descrition commit a09bb7f84befe432e7b3563bb213104a69e4f47e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:25:54 2023 +0100 fix transformations restrictions commit e723ac91ad3b3821248f2b00008d1ef91678a8dc Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 12:23:14 2023 +0100 refactor commit 1cba14e6ed2eb25c62a3987516d3cbdb989b4966 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 11:10:31 2023 +0100 rewrite backward commit d127fe60a7af3f3f99c5cca9b4f293e29fa39483 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 19:30:51 2023 +0100 forward sinking - added support 3d commit de447274aa81cd594b36842bff7838fd3d0131c8 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 17:07:16 2023 +0100 fix build errors unit tests commit b09cf32ccfbc1706c3107273819557adfe3adcef Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 14 14:29:01 2023 +0100 cleanup tests commit 8bb808756cc8d7f0571c35ccd78bda964d37f3c9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 14 14:16:15 2023 +0100 wrote unit tests draft commit b47571a3cd8a815ac4a84bb140d28304eed9629b Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 13 19:52:32 2023 +0100 implement transformation for 2d case commit 9a5fcd38b0162bbae73f49c482d1c125fdfe3181 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Feb 9 20:40:45 2023 +0100 added inital transformation commit 5e812990b9c8eaafa2b25a55607f8b1fb41f4eff Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 1 19:47:05 2023 +0100 add unit tests; fix commit 01dc4766ead37a455808a7b3b42a849c850144a5 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 1 13:57:34 2023 +0100 initial commit 2c24d97729d93dfc79d26d442e3e4f592ad0614c Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 13:36:35 2023 +0100 fix year commit 879f9e3d0ab115ef1a27e709c0aad4a5ce883f30 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 13:29:05 2023 +0100 add docs commit 030b2e4fdba826160fa45d16f9ba295c9bebc5ed Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:58:55 2023 +0100 cleanup commit 97ed5403ea2854064c5817edca11266106b06b95 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:52:49 2023 +0100 clang fixes + remove unneeded functions commit 4411e6ffbad5f63ce592025785127dda277fe70d Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:50:03 2023 +0100 fix unit tests execution commit 0483be0d6a31f18dfddfa1bca41ea75a833c0f03 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:33:03 2023 +0100 initial commit 114314b2a31112bc871b6f0c37568e4f177e9448 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 28 13:52:00 2023 +0100 initial commit d3493c48ae56723dead05d7135d177b9863b3229 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 19:38:35 2023 +0100 fix errors after rebase commit 141a77b3934beb912bcfc1aa19d2018cb11dab8f Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 28 13:27:52 2023 +0100 initial commit 69f8479260960c8d9e431ac1d4261264e205fc43 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 17:42:13 2023 +0100 Removed debug switch commit a4c43c40ee42c0a025eb9bef7a5ae933245f7c96 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:47:38 2023 +0100 Fixed unit tests build commit 506128f16c96d2fffe586e6095a98bb87033cf18 Merge: 4f2b13f34f 451772350c Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:41:01 2023 +0100 Merge branch 'feature/extract_gather_transpose' Enabled transformations to remove Transpose/Extract layers commit 4f2b13f34f8ff723226c82f9f8d13e389e989301 Merge: 8e5ca732de ba77c7d15b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:38:17 2023 +0100 Merge branch 'tests/bacward_compatibility' import/export tests commit 8e5ca732dec3b14a87976e8108e587e4acce0da2 Merge: cab1b42e38 919a7b67bf Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:35:35 2023 +0100 Merge branch 'cpu_execution' Preprocessing on CPU using ngraph commit ba77c7d15b Merge: 86e1b68d84 cf8dccaedb Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:27:46 2023 +0100 Merge branch 'master' into tests/bacward_compatibility commit 451772350c439561dfc4909b22fbafcd47f66011 Merge: c7cfb95728 cf8dccaedb Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:22:20 2023 +0100 Merge branch 'master' of https://github.com/mryzhov/openvino into feature/extract_gather_transpose commit cab1b42e38868c5ded27c76c3530917cea7dbae5 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 15:39:12 2023 +0100 fix utils StridedSlice to be sinkable commit 849f651f455ae8d75069b058de2f2419c30c9918 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Mar 6 15:24:10 2023 +0100 Ivan implementation commit 9fa205e9e60a8de36edba5b2e4ef05d48c56846a Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Mar 3 18:54:10 2023 +0100 initial commit e40497e9c100e3d0e11b6f03d928a8326fb1626b Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Mar 3 15:18:20 2023 +0100 add new transformations to pipeline commit c39c76bc19a1564816d8d9d1c586b3fdc09a8294 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 2 14:59:07 2023 +0100 add binary and gather_fuse commit ac37d4aa73f5491f9054c0e5228d01c477a980bc Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Mar 2 14:38:43 2023 +0100 initial commit 2d5383d415346d2c6ef460b502e2e53aed6be5f8 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 1 19:47:05 2023 +0100 add unit tests; fix commit 668da966b36d044fb53d29588f3aca3b63907e17 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Mar 1 13:57:34 2023 +0100 initial commit 3e881d1284f5fbf22219b0e2e935a0b72eaec886 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 28 13:52:00 2023 +0100 initial commit c067aec546a4997d7ad84fc4689796aad30d62a9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:45:19 2023 +0100 clang commit 4fb2f54fd0b86cc972396958019237f48cf19178 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:43:36 2023 +0100 remove debug commit 31776582f9e9934bdcfa3cd24381d90719ce6e10 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:41:39 2023 +0100 add descrition commit 20e390859b9757d1b35c7d3f16b06814a39248ce Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 13:25:54 2023 +0100 fix transformations restrictions commit b1cd65d75ff061f9e4ad69ea174fe37e321073d2 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 12:23:14 2023 +0100 refactor commit dab693e2b2c859c53ea99364b1721354729185f1 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Feb 22 11:10:31 2023 +0100 rewrite backward commit d33eafcce3ccdf72c46d33216ef1e80645e6a12e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 19:30:51 2023 +0100 forward sinking - added support 3d commit 8676c2a9ce707d370892d1514f1eb8309ee67da6 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 17:07:16 2023 +0100 fix build errors unit tests commit 030dab30945018ef6c91879d6c0a034a6e5d3783 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 14 14:29:01 2023 +0100 cleanup tests commit 5dfba630ac973a96b27aa294dc0e090d8fce8b4c Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 14 14:16:15 2023 +0100 wrote unit tests draft commit ae63a5aacc9467de3d74801dd61a26262998f686 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 13 19:52:32 2023 +0100 implement transformation for 2d case commit 23c19be01ea230f8e8ff655e58c6c0b9b8ac70ea Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Feb 9 20:40:45 2023 +0100 added inital transformation commit d69e88c2d2e8354bb3f3e11c24d1ab586eb12401 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 28 13:27:52 2023 +0100 initial commit 82e3f0ed4d0421aec686005a7a4e46de6f2499ac Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 13:36:35 2023 +0100 fix year commit b4eb780e588ff12f98277586f5c1da2a22597f55 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Feb 21 13:29:05 2023 +0100 add docs commit fd987b0652856af50463ed5952eec24478bb772b Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:58:55 2023 +0100 cleanup commit c852d051946312becae279e25bf4b972be6b1556 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:52:49 2023 +0100 clang fixes + remove unneeded functions commit aa3cc35161104fe568a3301cfe6bd7922f3e5294 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:50:03 2023 +0100 fix unit tests execution commit c4dd7246e13c154e6e1b3da72768cf794c54df9f Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Feb 20 18:33:03 2023 +0100 initial commit 919a7b67bfb43ff9716fde454cc7fa64308de8af Merge: 8df3c54eb8 cf8dccaedb Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:16:12 2023 +0100 Merge branch 'master' of https://github.com/mryzhov/openvino into cpu_execution commit c7cfb957280e39f3a13c93a0015cd912e594e98e Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 16:14:29 2023 +0100 clang formatted commit 6a9b1974992053318d122ddf7b2541069486a36e Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 15:53:44 2023 +0100 clang fixes commit 60cbd4d41267c9f2146bb1e13a56c05586e623dd Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 15:51:29 2023 +0100 Excluded gna supported trasposes commit b23bbc10a65ebcebcc0acc7298c124c7ddfb5a97 Merge: 084f147210 716f020956 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 11:47:34 2023 +0100 Merge branch 'feature/extract_gather_transpose' of https://github.com/mryzhov/openvino into feature/extract_gather_transpose commit 084f1472104e8a767483646856a54b50af4af408 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Mar 6 11:47:29 2023 +0100 Fixed double transposition commit 716f020956fe249672b0aea6c3bd30a081cfca6d Merge: 86ea8cb7dc 07f287e362 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 28 16:55:55 2023 +0100 Merge branch 'master' into feature/extract_gather_transpose commit 86ea8cb7dc2e88ffb40cf17dfa8de28af4144814 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 28 16:12:15 2023 +0100 clang fixes commit 38e69687943634430b77ebf2059771c1b13f5d96 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 28 12:23:07 2023 +0100 Added tests commit b5bda65d1b67c9d5a3d2a81cc6c8a0e3e108bc87 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 28 12:22:47 2023 +0100 fixed compile error commit 7815caa8e45d7498d9c56bcb2443a6ef6762da86 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 28 11:06:39 2023 +0100 Init commit 4e4da5b0097ecb055e1faf118931a148842a07e4 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Feb 23 10:24:57 2023 +0100 Gather transformation commit 8df3c54eb8a132f99e7a579269db57478ff738e3 Merge: e0e80f9dd8 98392a043b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Feb 22 17:04:55 2023 +0100 Merge branch 'master' into cpu_execution commit e0e80f9dd886204768a129eb426b1c6ec1a114c3 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Wed Feb 22 14:56:49 2023 +0100 WA I16->F16 commit 1d6319c257bd7d6d99e952286b86c3d8e05bcf00 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 15:50:29 2023 +0100 Reverted test commit 86e1b68d84 Merge: 9871d61042 94b64fed79 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 12:28:11 2023 +0100 Merge branch 'master' into tests/bacward_compatibility commit cd2c75cbd35975a8b3c93b8001dc38f6fab38f46 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 12:23:52 2023 +0100 fixed static build commit 290b592f4ace79d3bbff2906186481bfd5e267a0 Merge: d63cfcf774 94b64fed79 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 12:01:38 2023 +0100 Merge branch 'master' into cpu_execution commit d63cfcf7745652cfdfd38211ae3565153860a018 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 10:55:52 2023 +0100 clang format commit 71fedeef5d9c12b6bfe95e8ec3d2d6686d1619c6 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 10:30:25 2023 +0100 Revert "backward tests" This reverts commit 05dacd1a51a8756880e49523138af9b14245e262. commit 036fb95e79307c90e2f07add88683f47e5d0a857 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 20 16:07:52 2023 +0100 backward tests commit ed6b8d5b0da6f4dde4489681775cc5a26af0bc13 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 20 11:38:38 2023 +0100 reset gather_transformation.cpp to master commit 64c7f4e963e439151252d7d339baadd7d2722581 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 20 11:36:26 2023 +0100 reverted commit 9af3771cc31d61aa1eb09053c56625aa7e7c5927 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 20 11:31:59 2023 +0100 removed changes in gather commit d0a23572b5fad9176e58b816b0ab2ca0697119ff Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 20 11:28:48 2023 +0100 Revert "Added backward compatibility tests" This reverts commit afdd7afe72df88c96593f9f5b9bbfaaf222cec63. commit a504e700fcda8ada2a4753b71dd25b517951a47a Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 20 11:25:00 2023 +0100 Added trasposition check commit 560df30526beb7e5da57f8f2c42dbc4e63029da4 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Feb 16 18:32:25 2023 +0100 Removed Gather transformation commit 2c5b99d224df8acf39330483065e828de46af03b Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Feb 16 15:30:32 2023 +0100 Added backward compatibility tests commit e4935d02f5850b2f526412b0c30ee650f8175343 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 13 21:14:29 2023 +0100 compile warnings commit 86a36411027aa7aba97784c261ffcb64dd2ff92f Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 13 18:14:21 2023 +0100 clang formatting commit 308b239302c6ce0186cb95bfc31ccfd5122b1613 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Mon Feb 13 17:24:51 2023 +0100 Fixed merge issues commit 94818a4de96c068241b285858bb9317d20c13c94 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Feb 9 16:39:47 2023 +0100 passed all the tests commit fa5e3740e2d3f961ade7b6b894d11d23c549a4d6 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 7 13:02:21 2023 +0100 Fixed precison in gather transformation commit f0a70f054b251999d36b3291cd90a2c9e63e5f17 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 7 13:01:43 2023 +0100 Added the new export version commit cef818c7a2e0a22fdfd215576f60975070debebd Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 7 12:58:00 2023 +0100 Support of legacy transposes commit e71e476dd5c2feb2f699987d1dbade1271a45e88 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Thu Jan 26 16:47:41 2023 +0100 Pre_post_process init commit e1d297092c952fa9ee218b056632964a8941d2ec Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Thu Jan 19 14:52:02 2023 +0100 fix after rebase commit 45ab28de3aa6d6de42c9378a2952a1dbc23d3ba9 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Nov 23 19:10:24 2022 +0100 remove GatherIESubstitute transformation as not needed commit 4cb2b0aa3301ee68a0204cf3723a6465951626e1 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Wed Nov 23 19:08:57 2022 +0100 fix funcitional test commit 981d086f7f327f7b6665b6076da4d5aaff36be2e Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Tue Nov 22 14:23:47 2022 +0100 fix precision; left bug output EXACT commit 3cf71e0de1a47ac771d5611014d8e7e109168043 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Nov 18 18:45:12 2022 +0100 fixes commit c69d57f1f3b52dfbf6c68279353e4ca95eb26648 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Fri Nov 18 17:55:35 2022 +0100 fix functional test commit 1b540338910e2ad17f21ba7d71c098877dadfa03 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Nov 14 17:51:20 2022 +0100 add gather remove commit d7bff1ffa868b41f340cd17e66548d630b0822c2 Author: Evgeny Kotov <evgeny.kotov@intel.com> Date: Mon Nov 14 16:11:44 2022 +0100 initial commit 9871d61042 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 11:10:38 2023 +0100 Fixed build of deprecated tests commit cbf5246c1a Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 10:46:15 2023 +0100 clang format commit c34f81b899 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Tue Feb 21 10:28:33 2023 +0100 Migrated the deprecated backward compatibility tests to the new infra commit bc1dd9d302 Author: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Date: Sun Mar 19 18:15:21 2023 +0100 Squashed commit of the following: commit a56a0768f1 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:48:23 2023 +0000 remove TransposeSinking from MOC commit 430bdbec35 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:39:22 2023 +0000 delete debug serialize commit a40658e369 Merge: 346796af9c cf70ced2a9 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:33:46 2023 +0000 Merge branch 'itikhono/ts/slice' of https://github.com/itikhono/openvino into itikhono/ts/slice commit 346796af9c Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:33:27 2023 +0000 fix TSSqueeze/TSUnsqueeze transformations commit cf70ced2a9 Merge: 7a6988a4a6 a9bd5f741d Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Fri Mar 17 13:12:18 2023 +0400 Merge branch 'master' into itikhono/ts/slice commit 7a6988a4a6 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 08:50:15 2023 +0000 codestyle commit 849dc70763 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 16:32:45 2023 +0400 refactoring commit 958f000e02 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 16:20:06 2023 +0400 Added TSSlice transformation to TSGeneral, created TransposeSinkingGeneral alias in ov::pass namespace commit 83ab2cc5f6 Merge: 1f8fdfb353 88ddbb2437 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 16:09:35 2023 +0400 Merge branch 'itikhono/ts/slice' of https://github.com/itikhono/openvino into itikhono/ts/slice commit 1f8fdfb353 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 16:06:15 2023 +0400 fix warnings commit 66d16ae45e Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 00:07:44 2023 +0400 fix minor mistakes commit 1ac7690d42 Merge: f1dc3702f1 6ffa8da922 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 00:02:39 2023 +0400 Merge remote-tracking branch 'upstream/master' into itikhono/ts/slice commit f1dc3702f1 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 16 23:48:10 2023 +0400 TSReduction refactoring, move Unsqueeze/Squeeze transformations to separate files, added limited support for Reshape op + tests commit 88ddbb2437 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 16 16:16:59 2023 +0000 codestyle commit 2bc1334f65 Merge: ce84ce24ec 4b9e048e18 Author: Ivan <ivan.tikhonov@intel.com> Date: Wed Mar 15 17:10:30 2023 +0400 Merge branch 'itikhono/ts/refactoring' into itikhono/ts/slice commit ce84ce24ec Merge: 6e6361f83a 36c18e29a8 Author: Ivan <ivan.tikhonov@intel.com> Date: Wed Mar 15 16:25:41 2023 +0400 Merge remote-tracking branch 'upstream/master' into itikhono/ts/slice commit 4b9e048e18 Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 22:54:33 2023 +0400 resolve merge conflicts commit 169a722212 Merge: ee42aaf13b 596036a2db Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 22:48:22 2023 +0400 Merge remote-tracking branch 'upstream/master' into itikhono/ts/refactoring commit ee42aaf13b Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 14:43:50 2023 +0000 codestyle commit b3342f7c46 Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 22:30:09 2023 +0400 TransposeSinking refactoring: class names, namespaces commit e7759e1fa0 Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 21:59:28 2023 +0400 TransposeSinking Rafactoring part2: move the transformations to a separate folder, align namespaces commit 7b91ac1e6b Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 17:59:12 2023 +0400 fix a warning commit d284ac1b7a Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 17:48:46 2023 +0400 Add descriptions to the transformations, add additional checks commit 6e6361f83a Merge: 54bf0444e4 3a96e06d4c Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 14 16:20:34 2023 +0400 Merge remote-tracking branch 'upstream/master' into itikhono/ts/slice commit 54bf0444e4 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 10 19:54:19 2023 +0400 TransposeSinking: add support for Slice op commit a47a18cf55 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 9 22:29:37 2023 +0400 Add new tests for Unsqueeze/Squeeze; refactoring; remove debug code commit e13d0e589f Merge: fa8e81f0f1 3d52fc843a Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Thu Mar 9 01:43:47 2023 +0400 Merge branch 'master' into itikhono/ts/fix_performance_issues commit fa8e81f0f1 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 9 04:20:08 2023 +0400 Fix warning as error on Windows commit c77e98caf6 Merge: 99192feb02 3eeaf7f9bd Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 9 02:42:47 2023 +0400 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit 99192feb02 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 9 02:42:37 2023 +0400 fix unary tests commit 11b500953d Merge: b60015f90d 3b924ca206 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 9 01:33:57 2023 +0400 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit b60015f90d Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 9 01:33:46 2023 +0400 Add tests for TS backward transformations, update TransposeSinkingFuse transformation, delete StridedSlice transformation prototype + tests refactoring commit 3eeaf7f9bd Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Wed Mar 8 17:42:25 2023 +0000 codestyle commit db09fe4965 Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 7 16:55:40 2023 +0400 fix build commit 3b924ca206 Merge: 54b4061282 db09fe4965 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 7 08:56:32 2023 +0000 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit 54b4061282 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 7 08:51:48 2023 +0000 TransposeReduction: skip the case when 2nd input for Squeeze is not provided commit 176686318f Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 7 07:34:40 2023 +0000 codestye commit 3c5f62c013 Merge: ef0e89551d 0dad7749b5 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 7 07:33:09 2023 +0000 Merge remote-tracking branch 'upstream/master' into itikhono/ts/fix_performance_issues commit ef0e89551d Author: Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 7 15:30:22 2023 +0400 Transpose sinking tests refactoring: part 3. + Revert changes in MOC. commit d71949fd09 Author: Ivan <ivan.tikhonov@intel.com> Date: Mon Mar 6 18:23:40 2023 +0400 Add limited support for StridedSlice op commit 3565ff2181 Author: Ivan <ivan.tikhonov@intel.com> Date: Mon Mar 6 17:45:52 2023 +0400 TransposeSinking tests refactoring: part2 commit 34c89eb962 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 3 04:34:48 2023 +0400 TransposeSinking tests refactoring: part1 commit c5991f0b06 Merge: 6de1336894 caea77c132 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 17:21:17 2023 +0000 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit caea77c132 Merge: 199114a4c6 873d15ce1a Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 21:39:57 2023 +0400 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit 199114a4c6 Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 21:39:12 2023 +0400 add include commit 6de1336894 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 17:21:02 2023 +0000 add the missed line commit 123835c86d Author: Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 19:09:36 2023 +0400 Support TS for Interpolate, VariadicSplit, IsInf, IsNan, IsFinite + refactoring commit 873d15ce1a Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 13:33:40 2023 +0000 fix gna build commit 13f17d254b Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Mar 2 12:24:12 2023 +0000 fix TS for Interpolate + codestyle commit b769d21912 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Feb 23 17:14:30 2023 +0000 Add SpaceToBatch/BatchToSpace commit 9d8016d1e6 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Feb 23 17:12:54 2023 +0000 Fix validate for split, revert changes for concat, add BatchToSpace/SpaceToBatch commit 20579455b7 Merge: 3baf0c7900 d9fc5bac80 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Feb 23 10:32:14 2023 +0000 Merge remote-tracking branch 'upstream/master' into itikhono/ts/fix_performance_issues commit 3baf0c7900 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Thu Feb 23 10:31:56 2023 +0000 fix node validation commit fa9fe34c16 Merge: ef6e141082 eaf368a5f5 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Wed Feb 22 15:33:52 2023 +0000 Merge remote-tracking branch 'upstream/master' into itikhono/ts/fix_performance_issues commit ef6e141082 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Feb 21 14:34:26 2023 +0000 Add additional validations commit 81b9e6eece Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Mon Feb 20 13:54:27 2023 +0000 delete debug print commit 9333c1cac5 Merge: aa1aadd3c4 08293d39e1 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Mon Feb 20 13:51:45 2023 +0000 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit aa1aadd3c4 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Mon Feb 20 13:50:42 2023 +0000 Fix TransposeReduction, fix TransposeSinkingSplit, add unsqueeze support commit 20168b251a Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Feb 17 07:46:12 2023 +0000 fix TransposeReduce transformations commit 08293d39e1 Merge: 43ef82320f 5cc1c2c5e5 Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Tue Feb 14 18:11:20 2023 +0400 Merge branch 'master' into itikhono/ts/fix_performance_issues commit 43ef82320f Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Feb 14 14:09:25 2023 +0000 fix the issue in TransposeFuse transformation commit 6e3fcf74e2 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Feb 14 08:09:25 2023 +0000 Fix TransposeReduceBackward commit a34e3ced9e Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Mon Feb 13 17:08:19 2023 +0000 fix ts for Concat and Reduce commit 8598d64174 Merge: d9ea97bf4b f5bff5c087 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Mon Feb 13 07:46:32 2023 +0000 Merge branch 'itikhono/ts/fix_performance_issues' of https://github.com/itikhono/openvino into itikhono/ts/fix_performance_issues commit d9ea97bf4b Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Mon Feb 13 07:45:54 2023 +0000 fix warning as error, fix tests failures commit f5bff5c087 Merge: 1de806f9f7 b300df1be6 Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Sun Feb 12 21:51:00 2023 +0400 Merge branch 'master' into itikhono/ts/fix_performance_issues commit 1de806f9f7 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Sun Feb 12 17:48:05 2023 +0000 codestyle commit 2c48d6c583 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Sun Feb 12 17:47:32 2023 +0000 Resolve the performance issues in TransposeSinking transformation commit db220f11ee Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 09:42:42 2023 +0000 delete debug code commit 8f8f0e821b Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Tue Mar 21 09:41:48 2023 +0000 fix TSSqueeze/TSUnsqueeze transformations in case of Reshape op commit 981e8ad3c0 Merge: a56a0768f1 c5f65eea73 Author: Ivan <ivan.tikhonov@intel.com> Date: Mon Mar 20 19:47:18 2023 +0400 Merge remote-tracking branch 'upstream/master' into itikhono/ts/slice commit a56a0768f1 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:48:23 2023 +0000 remove TransposeSinking from MOC commit 430bdbec35 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:39:22 2023 +0000 delete debug serialize commit a40658e369 Merge: 346796af9c cf70ced2a9 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:33:46 2023 +0000 Merge branch 'itikhono/ts/slice' of https://github.com/itikhono/openvino into itikhono/ts/slice commit 346796af9c Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 15:33:27 2023 +0000 fix TSSqueeze/TSUnsqueeze transformations commit cf70ced2a9 Merge: 7a6988a4a6 a9bd5f741d Author: Ivan Tikhonov <ivan.tikhonov@intel.com> Date: Fri Mar 17 13:12:18 2023 +0400 Merge branch 'master' into itikhono/ts/slice commit 7a6988a4a6 Author: Tikhonov Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 08:50:15 2023 +0000 codestyle commit 849dc70763 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 16:32:45 2023 +0400 refactoring commit 958f000e02 Author: Ivan <ivan.tikhonov@intel.com> Date: Fri Mar 17 16:20:06 2023 +0400 Added TSSlice transformation to TSG… * Merge fixes * clang fixes * fix Reshape is unsupported; Transpose backward sinking case * fix functional test params * fix bug : duplicated source code in GNAGraphCompiler::PoolingPrimitive after rebase * remove legacy transformations * Fixed out of range issue * aligned memory for convolution with padding # Conflicts: # src/plugins/intel_gna/tests/functional/shared_tests_instances/subgraph_tests/const_conv_concat.cpp * clang fixes * [GNA] Limitations refactoring (#16957) * Limitations refactoring * fix CI builds/tests * changes after review * Move GraphCompiler initialization to constructor * resolve conflicts after rebase * update after review * resolve problem with double initialization for Limitations # Conflicts: # src/plugins/intel_gna/src/backend/gna_limitations.cpp # src/plugins/intel_gna/src/backend/gna_limitations.hpp # src/plugins/intel_gna/src/gna_transformations_pipeline.cpp * Rebase fixes * Removed duplicate of is_transpose_supported * Rebase fix 2 * Reduced code duplication * Clang code style * fix unit tests * Removed legacy failed tests * add gather_sinking_test * Corrected tests to produce unsupported transposes * Reverted unnecessary fix * hide deprecated warnings * clang fix * fixed concat-gather transformation * revert back ts_squeeze bugfix * remove workaround * fix build * remove concat_transpose.cpp * remove backward_compatibility.cpp test * remove debug * fix std::equal * Revert "remove debug" This reverts commit 574a3a04f3. * remove debug * return concat_transpose with fix * rotates convolution inputs * fix after rebase * cleanup after rebase * fixes after rebase * fixes after rebase * fix after rebase GatherSinkingGeneralBackward - return back GatherSinkingSplitBackward * fix after rebase GatherSinkingGeneralBackward - return back * clang fix * windows build fix; fix rebase issue * code review fixes * clang fixes * code review fixes * code review fixes * code review fixes - rename ts_split and ts_concat files * code review fixes : move auxilary functions from pipeline to transformation helper * code review fixes -> move has_one_consumer to graph_utils * fix comparing * cleanup * code review fix - graph_compiler remove unneeded copying * code review fixes - remove unneeded code * code review fix gna_layer_info * code review fix - remove unneeded code from LoadNetwork * cleanup * update modules * code review fixes * code review fixes * code review fixes * Revert "code review fixes -> move has_one_consumer to graph_utils" This reverts commit f1a24041083b0da3c5ca34417e1789fd54e24dab. * Revert "code review fixes" This reverts commit d01b58a5b1a7a70bf16eec6f86b8dc1e50bf65c7. * revert commits * code review fixes * code review fixes - remove unneeded * code review fixes * code review fix --------- Co-authored-by: Mikhail Ryzhov <mikhail.ryzhov@intel.com> Co-authored-by: Tomasz Adamowicz <tomasz.adamowicz@intel.com>
2023-06-28 19:55:34 +02:00
parent 18e737493c
commit 7ae8826b52
42 changed files with 1149 additions and 542 deletions
--- a/src/plugins/intel_gna/src/backend/gna_limitations.cpp
+++ b/src/plugins/intel_gna/src/backend/gna_limitations.cpp
@@ -18,6 +18,7 @@
 #include "gna_lib_ver_selector.hpp"
 #include "ie_ngraph_utils.hpp"
 #include "log/log.hpp"
+#include "openvino/opsets/opset12.hpp"

 namespace std {
 inline std::ostream& operator<<(std::ostream& os, const std::set<ov::element::Type>& t) {
@@ -35,6 +36,7 @@ inline std::ostream& operator<<(std::ostream& os, const std::set<ov::element::Ty
 namespace ov {
 namespace intel_gna {
 using namespace target;
+using namespace opset12;
 namespace limitations {

 class SupportedElementTypes {
@@ -689,22 +691,6 @@ void Limitations::init(const DeviceVersion& compile_target) {
    k_instance = std::shared_ptr<Limitations>(new Limitations(compile_target));
 }

-bool Limitations::is_transpose_2d(const std::vector<size_t>& shape) {
-    return std::count_if(std::begin(shape), std::end(shape), [](size_t dim) {
-               return dim != 1;
-           }) == 2;
-}
-
-bool Limitations::is_transpose_supported(const std::vector<size_t>& shape) {
-    if (!is_transpose_2d(shape))
-        return false;
-    auto shape_no_1 = shape;
-    shape_no_1.erase(std::remove(shape_no_1.begin(), shape_no_1.end(), 1), shape_no_1.end());
-    size_t min, max;
-    std::tie(min, max) = std::minmax(shape_no_1[0], shape_no_1[1]);
-    return min <= 8 && max % 8 == 0 && max >= 8 && max <= kTransposeMaxSize;
-}
-
 size_t Limitations::get_min_batch_to_fit_in_buffer(InferenceEngine::DataPtr input) {
    auto total_size = InferenceEngine::details::product(std::begin(input->getDims()), std::end(input->getDims()));
    return total_size / kBufferMaxSize + 1;
@@ -753,31 +739,38 @@ bool SupportedElementTypes::IsConstantTypeSupported(ov::element::Type elem_type,
    return true;
 }

-bool Limitations::is_transpose_supported(const std::shared_ptr<const ov::Node>& node) {
-    OPENVINO_ASSERT(node, "Transpose node is empty!");
-    const ov::Shape squeezed_shape = graph_utils::squeeze_shape(node->get_input_shape(0));
-    const size_t min_input_dim = std::min(squeezed_shape[0], squeezed_shape[1]);
-    const size_t max_input_dim = std::max(squeezed_shape[0], squeezed_shape[1]);
+bool Limitations::is_transpose_supported(const ov::Shape& shape) {
+    const ov::Shape squeezed_shape = graph_utils::squeeze_shape(shape);

    // GNA transpose limitations:
    // - supports 2d transposes only
    // - smaller dimension should be less or equal to 8
-    // - bigger dimension should be a multiple of Limitations::kNoOfInputsDivisor
-    if (squeezed_shape.size() == 2 && min_input_dim <= 8 && ALIGN(max_input_dim, kNoOfInputsDivisor) == max_input_dim) {
-        return true;
+    // - bigger dimension should be a multiple of limitations::noOfInputsDivisor
+    if (squeezed_shape.size() == 2) {
+        const size_t min_input_dim = std::min(squeezed_shape[0], squeezed_shape[1]);
+        const size_t max_input_dim = std::max(squeezed_shape[0], squeezed_shape[1]);
+        if (min_input_dim <= 8 && max_input_dim % Limitations::kNoOfInputsDivisor == 0 &&
+            max_input_dim <= kTransposeMaxSize) {
+            return true;
+        }
    }
    return false;
 }

-bool Limitations::is_conv_supported(const std::shared_ptr<ngraph::op::ConvolutionIE>& conv_ie,
+bool Limitations::is_transpose_supported(const std::shared_ptr<const ov::Node>& node) {
+    OPENVINO_ASSERT(node, "Transpose node is empty!");
+    return is_transpose_supported(node->get_input_shape(0));
+}
+
+bool Limitations::is_conv_supported(const std::shared_ptr<ov::intel_gna::op::GNAConvolution>& conv_gna,
                                    const InferenceEngine::Precision gna_precision,
                                    bool is_exception_allowed) {
-    OPENVINO_ASSERT(conv_ie, "ConvolutionIE node is empty!");
-    size_t batch_size = conv_ie->input_value(0).get_shape()[0];
+    OPENVINO_ASSERT(conv_gna, "GNAConvolution node is empty!");
+    size_t batch_size = conv_gna->input_value(0).get_shape()[0];
    if (batch_size != 1) {
        if (is_exception_allowed) {
-            THROW_GNA_EXCEPTION << "topology with layer: " + conv_ie->get_friendly_name() +
-                                       ", type: " + conv_ie->get_type_name() + ", and batch size(" +
+            THROW_GNA_EXCEPTION << "topology with layer: " + conv_gna->get_friendly_name() +
+                                       ", type: " + conv_gna->get_type_name() + ", and batch size(" +
                                       std::to_string(batch_size) + ") != 1 not supported";
        }
        return false;
@@ -789,15 +782,15 @@ bool Limitations::is_conv_supported(const std::shared_ptr<ngraph::op::Convolutio
                                                           static_cast<uint32_t>(filter_stride_width));
        return cnn2d::AbstractValidator::ValidationSuccesful(is_exception_allowed,
                                                             error,
-                                                             conv_ie->get_friendly_name(),
-                                                             conv_ie->get_type_name());
+                                                             conv_gna->get_friendly_name(),
+                                                             conv_gna->get_type_name());
    };
-    auto input_shape = conv_ie->input_value(0).get_shape();
-    auto filter_shape = conv_ie->input_value(1).get_shape();
-    if ((4 == filter_shape.size() && filter_shape[2] > 1 && filter_shape[3] > 1) ||
-        (4 == input_shape.size() && input_shape[2] > 1 && input_shape[3] > 1)) {
+    auto input_shape = conv_gna->input_value(0).get_shape();
+    auto filter_shape = conv_gna->input_value(1).get_shape();
+    if ((4 == filter_shape.size() && filter_shape[1] > 1 && filter_shape[2] > 1) ||
+        (4 == input_shape.size() && input_shape[1] > 1 && input_shape[2] > 1)) {
        pass::helper::ConvData conv_data;
-        pass::helper::GetConvData(conv_ie, conv_data);
+        pass::helper::GetConvData(conv_gna, conv_data);
        if (gna_convolution_layer::isMappableFrom2DTo1D(static_cast<uint32_t>(conv_data.input_height),
                                                        static_cast<uint32_t>(conv_data.input_width),
                                                        static_cast<uint32_t>(conv_data.input_channel_count),
@@ -809,7 +802,7 @@ bool Limitations::is_conv_supported(const std::shared_ptr<ngraph::op::Convolutio
        }

        if (m_cnn_validator) {
-            return m_cnn_validator->ValidateCnn2D(conv_ie->get_friendly_name(),
+            return m_cnn_validator->ValidateCnn2D(conv_gna->get_friendly_name(),
                                                  static_cast<uint32_t>(conv_data.input_height),
                                                  static_cast<uint32_t>(conv_data.input_width),
                                                  static_cast<uint32_t>(conv_data.input_channel_count),
@@ -824,10 +817,12 @@ bool Limitations::is_conv_supported(const std::shared_ptr<ngraph::op::Convolutio
                                                  is_exception_allowed);
        }
    }
-    return check_dilation(conv_ie->get_dilations()[0], conv_ie->get_dilations()[1]);
+
+    return check_dilation(conv_gna->get_dilations()[0],
+                          conv_gna->get_dilations()[conv_gna->get_dilations().size() - 1]);
 }

-bool Limitations::is_pooling_supported(const std::shared_ptr<ngraph::opset7::MaxPool> max_pool,
+bool Limitations::is_pooling_supported(const std::shared_ptr<ov::intel_gna::op::GNAMaxPool> max_pool,
                                       bool is_exception_allowed) {
    OPENVINO_ASSERT(max_pool, "MaxPool node is empty!");
    auto kernels = max_pool->get_kernel();
@@ -869,6 +864,100 @@ bool Limitations::is_split_supported(const std::shared_ptr<ov::Node>& node, bool
    return is_aligned;
 }

+bool Limitations::is_concat_supported(const std::shared_ptr<const ov::Node>& node) {
+    OPENVINO_ASSERT(node, "Concat node is empty!");
+    auto concat_node = std::dynamic_pointer_cast<const Concat>(node);
+    const ov::Shape& output_shape = concat_node->get_output_shape(0);
+    auto axis = concat_node->get_axis();
+
+    return graph_utils::get_first_valuable_dim_id(output_shape) == axis;
+}
+
+bool Limitations::is_forward_transposed_concat_supported(const std::shared_ptr<const ov::Node>& node,
+                                                         const AxisVector& order) {
+    auto concat_node = std::dynamic_pointer_cast<const Concat>(node);
+    if (!concat_node) {
+        log::debug() << "Concat node is empty!" << std::endl;
+        return false;
+    }
+
+    const ov::Shape& output_shape = concat_node->get_output_shape(0);
+    auto axis = concat_node->get_axis();
+
+    const ov::Shape& transposed_shape =
+        graph_utils::transpose_shape(output_shape, pass::helper::reverse_transpose_order(order));
+    const size_t transposed_concat_axis = order[axis];
+
+    return graph_utils::get_first_valuable_dim_id(transposed_shape) == static_cast<int64_t>(transposed_concat_axis);
+}
+
+bool Limitations::is_backward_transposed_concat_supported(const std::shared_ptr<const ov::Node>& node,
+                                                          const AxisVector& order) {
+    auto concat_node = std::dynamic_pointer_cast<const Concat>(node);
+    if (!concat_node) {
+        log::debug() << "Concat node is empty!" << std::endl;
+        return false;
+    }
+
+    const ov::Shape& output_shape = concat_node->get_output_shape(0);
+    auto axis = concat_node->get_axis();
+
+    const ov::Shape& transposed_shape = graph_utils::transpose_shape(output_shape, order);
+    const size_t transposed_concat_axis = order[axis];
+
+    return graph_utils::get_first_valuable_dim_id(transposed_shape) == static_cast<int64_t>(transposed_concat_axis);
+}
+
+bool Limitations::is_forward_transposed_split_supported(const std::shared_ptr<const ov::Node>& node,
+                                                        const AxisVector& order) {
+    std::shared_ptr<const ov::Node> split_node = nullptr;
+    if (std::dynamic_pointer_cast<const Split>(node)) {
+        split_node = std::dynamic_pointer_cast<const Split>(node);
+    } else if (std::dynamic_pointer_cast<const VariadicSplit>(node)) {
+        split_node = std::dynamic_pointer_cast<const VariadicSplit>(node);
+    } else {
+        log::debug() << "Split node is empty!" << std::endl;
+        return false;
+    }
+
+    const ov::Shape& output_shape = split_node->get_output_shape(0);
+    auto constant_node = as_type_ptr<Constant>(split_node->input_value(1).get_node_shared_ptr());
+    if (!constant_node)
+        return false;
+    auto axis = constant_node->get_axis_vector_val()[0];
+
+    const ov::Shape& transposed_shape =
+        graph_utils::transpose_shape(output_shape, pass::helper::reverse_transpose_order(order));
+    const size_t transposed_concat_axis = order[axis];
+
+    return graph_utils::get_first_valuable_dim_id(transposed_shape) == static_cast<int64_t>(transposed_concat_axis);
+}
+
+bool Limitations::is_backward_transposed_split_supported(const std::shared_ptr<const ov::Node>& node,
+                                                         const AxisVector& order) {
+    std::shared_ptr<const ov::Node> split_node = nullptr;
+    if (std::dynamic_pointer_cast<const Split>(node)) {
+        split_node = std::dynamic_pointer_cast<const Split>(node);
+    } else if (std::dynamic_pointer_cast<const VariadicSplit>(node)) {
+        split_node = std::dynamic_pointer_cast<const VariadicSplit>(node);
+    } else {
+        log::debug() << "Split node is empty!" << std::endl;
+        return false;
+    }
+
+    const ov::Shape& output_shape = split_node->get_output_shape(0);
+    auto constant_node = as_type_ptr<Constant>(split_node->input_value(1).get_node_shared_ptr());
+    if (!constant_node)
+        return false;
+    auto axis = constant_node->get_axis_vector_val()[0];
+
+    const ov::Shape& transposed_shape =
+        graph_utils::transpose_shape(output_shape, pass::helper::reverse_transpose_order(order));
+    const int64_t transposed_concat_axis = order[axis];
+
+    return graph_utils::get_first_valuable_dim_id(transposed_shape) == transposed_concat_axis;
+}
+
 bool Limitations::is_op_supported(const std::shared_ptr<ov::Node>& node,
                                  const InferenceEngine::Precision gna_precision,
                                  bool is_exception_allowed) {
@@ -876,12 +965,13 @@ bool Limitations::is_op_supported(const std::shared_ptr<ov::Node>& node,
        return SupportedElementTypes::IsParameterTypeSupported(node->get_element_type(), is_exception_allowed);
    } else if (ov::op::util::is_constant(node)) {
        return SupportedElementTypes::IsConstantTypeSupported(node->get_element_type(), is_exception_allowed);
-    } else if (auto conv_ie = std::dynamic_pointer_cast<ngraph::op::ConvolutionIE>(node)) {
-        return is_conv_supported(conv_ie, gna_precision, is_exception_allowed);
+    } else if (auto conv = std::dynamic_pointer_cast<ov::intel_gna::op::GNAConvolution>(node)) {
+        return is_conv_supported(conv, gna_precision, is_exception_allowed);
    } else if (auto fully_connected = std::dynamic_pointer_cast<ngraph::op::FullyConnected>(node)) {
        return is_fc_supported(fully_connected, is_exception_allowed);
    } else if (ov::intel_gna::graph_utils::is_pooling(node)) {
-        return is_pooling_supported(std::dynamic_pointer_cast<ngraph::opset7::MaxPool>(node), is_exception_allowed);
+        return is_pooling_supported(std::dynamic_pointer_cast<ov::intel_gna::op::GNAMaxPool>(node),
+                                    is_exception_allowed);
    } else if (ov::op::util::is_output(node) || ov::op::util::is_sink(node) ||
               ov::intel_gna::graph_utils::is_eltwise_add(node) || ov::intel_gna::graph_utils::is_eltwise_mul(node) ||
               ov::intel_gna::graph_utils::is_crop_affined(node) ||
@@ -891,11 +981,11 @@ bool Limitations::is_op_supported(const std::shared_ptr<ov::Node>& node,
               (std::dynamic_pointer_cast<ov::op::util::ReadValueBase>(node) != nullptr) ||
               (std::dynamic_pointer_cast<ngraph::op::ScaleShiftIE>(node) != nullptr) ||
               (std::dynamic_pointer_cast<ngraph::op::PowerIE>(node) != nullptr) ||
-               (std::dynamic_pointer_cast<ngraph::opset9::MatMul>(node) != nullptr)) {
+               (std::dynamic_pointer_cast<MatMul>(node) != nullptr)) {
        return true;
    } else if (ov::intel_gna::graph_utils::is_gna_precision_agnostic(node)) {
-        if ((std::dynamic_pointer_cast<ngraph::opset9::Split>(node) != nullptr) ||
-            (std::dynamic_pointer_cast<ngraph::opset9::VariadicSplit>(node) != nullptr)) {
+        if ((std::dynamic_pointer_cast<Split>(node) != nullptr) ||
+            (std::dynamic_pointer_cast<VariadicSplit>(node) != nullptr)) {
            return is_split_supported(node, is_exception_allowed);
        }
        // TODO check concat are aligned when transformation will be moved to ngraph
--- a/src/plugins/intel_gna/src/backend/gna_limitations.hpp
+++ b/src/plugins/intel_gna/src/backend/gna_limitations.hpp
@@ -20,6 +20,8 @@
 #include "legacy/ngraph_ops/fully_connected.hpp"
 #include "ngraph/opsets/opset7.hpp"
 #include "ngraph/opsets/opset9.hpp"
+#include "ops/gna_convolution.hpp"
+#include "ops/gna_max_pool.hpp"

 namespace ov {
 namespace intel_gna {
@@ -173,8 +175,6 @@ public:
     */
    static inline std::shared_ptr<Limitations> get_instance();

-    static bool is_transpose_2d(const std::vector<size_t>& shape);
-    static bool is_transpose_supported(const std::vector<size_t>& shape);
    static size_t get_min_batch_to_fit_in_buffer(InferenceEngine::DataPtr input);

    /**
@@ -202,6 +202,13 @@ public:
     * @return true if supported
     */
    static bool is_split_supported(const std::shared_ptr<ov::Node>& node, bool is_exception_allowed = false);
+
+    /**
+     * @brief Validates if transpose is supported by GNA
+     * @param shape transpose
+     * @return true if supported
+     */
+    static bool is_transpose_supported(const ov::Shape& shape);
    /**
     * @brief Validates if transpose is supported by GNA
     * @param node transpose
@@ -209,13 +216,13 @@ public:
     */
    static bool is_transpose_supported(const std::shared_ptr<const ov::Node>& node);
    /**
-     * @brief Validates if legacy convolution is supported by GNA
-     * @param conv_ie convolution
+     * @brief Validates if convolution is supported by GNA
+     * @param conv_gna GNA convolution
     * @param gna_precision GNA inference precision
     * @param is_exception_allowed flag specifies whether exception is allowed
     * @return true if supported
     */
-    bool is_conv_supported(const std::shared_ptr<ngraph::op::ConvolutionIE>& conv_ie,
+    bool is_conv_supported(const std::shared_ptr<ov::intel_gna::op::GNAConvolution>& conv_gna,
                           const InferenceEngine::Precision gna_precision,
                           bool is_exception_allowed = false);
    /**
@@ -224,9 +231,19 @@ public:
     * @param is_exception_allowed flag specifies whether exception is allowed
     * @return true if precision is found in supported
     */
-    bool is_pooling_supported(const std::shared_ptr<ngraph::opset7::MaxPool> max_pool,
+    bool is_pooling_supported(const std::shared_ptr<ov::intel_gna::op::GNAMaxPool> max_pool,
                              bool is_exception_allowed = false);

+    static bool is_concat_supported(const std::shared_ptr<const ov::Node>& node);
+    static bool is_forward_transposed_concat_supported(const std::shared_ptr<const ov::Node>& node,
+                                                       const AxisVector& order);
+    static bool is_backward_transposed_concat_supported(const std::shared_ptr<const ov::Node>& node,
+                                                        const AxisVector& order);
+    static bool is_forward_transposed_split_supported(const std::shared_ptr<const ov::Node>& node,
+                                                      const AxisVector& order);
+    static bool is_backward_transposed_split_supported(const std::shared_ptr<const ov::Node>& node,
+                                                       const AxisVector& order);
+
    /**
     * @brief Validates if operation is supported by GNA
     * @param node operation
--- a/src/plugins/intel_gna/src/common/graph_utils.hpp
+++ b/src/plugins/intel_gna/src/common/graph_utils.hpp
@@ -198,7 +198,8 @@ inline bool is_eltwise_add(const std::shared_ptr<ngraph::Node>& node) {
 }

 inline bool is_pooling(const std::shared_ptr<ngraph::Node>& node) {
-    return (std::dynamic_pointer_cast<ngraph::opset7::MaxPool>(node) != nullptr);
+    return ((std::dynamic_pointer_cast<ngraph::opset7::MaxPool>(node) != nullptr) ||
+            std::dynamic_pointer_cast<ov::intel_gna::op::GNAMaxPool>(node) != nullptr);
 }

 template <typename T>
@@ -268,7 +269,7 @@ inline bool has_32bit_output(const std::shared_ptr<ngraph::Node>& node) {
    return ((std::dynamic_pointer_cast<ngraph::op::FullyConnected>(node) != nullptr) ||
            (std::dynamic_pointer_cast<ngraph::opset9::MatMul>(node) != nullptr) ||
            (std::dynamic_pointer_cast<ngraph::opset9::Convolution>(node) != nullptr) ||
-            (std::dynamic_pointer_cast<ngraph::op::ConvolutionIE>(node) != nullptr) ||
+            (std::dynamic_pointer_cast<ov::intel_gna::op::GNAConvolution>(node) != nullptr) ||
            (std::dynamic_pointer_cast<ngraph::opset9::Add>(node) != nullptr) ||
            (std::dynamic_pointer_cast<ngraph::opset9::Multiply>(node) != nullptr) ||
            (std::dynamic_pointer_cast<ngraph::op::Eltwise>(node) != nullptr) ||
@@ -625,6 +626,20 @@ bool has_child_node(std::shared_ptr<ov::Node> node) {
    return false;
 }

+/**
+ * @brief Checks if shape without dimensions == 1 is 2D
+ */
+inline bool is_shape_2d(const ov::Shape& shape) {
+    return graph_utils::squeeze_shape(shape).size() == 2;
+}
+
+/**
+ * @brief Checks if node has N consumers
+ */
+inline bool has_n_consumers(const std::shared_ptr<ov::Node>& node, size_t n_consumers) {
+    return node->output(0).get_target_inputs().size() == n_consumers;
+}
+
 }  // namespace graph_utils
 }  // namespace intel_gna
 }  // namespace ov
--- a/src/plugins/intel_gna/src/descriptions/gna_desc.hpp
+++ b/src/plugins/intel_gna/src/descriptions/gna_desc.hpp
@@ -74,9 +74,11 @@ struct GnaDesc {
    }

    InferenceEngine::DataPtr to_ie_data() {
+        OPENVINO_SUPPRESS_DEPRECATED_START
        return std::make_shared<InferenceEngine::Data>(
            name,
            InferenceEngine::TensorDesc(model_precision, dims, model_layout));
+        OPENVINO_SUPPRESS_DEPRECATED_END
    }
 };

@@ -98,9 +100,11 @@ struct InputDesc : GnaDesc {
    }

    InferenceEngine::InputInfo::Ptr ToIEInputInfo() {
+        OPENVINO_SUPPRESS_DEPRECATED_START
        InferenceEngine::InputInfo::Ptr input_info = std::make_shared<InferenceEngine::InputInfo>();
        input_info->setInputData(this->to_ie_data());
        return input_info;
+        OPENVINO_SUPPRESS_DEPRECATED_END
    }
 };

--- a/src/plugins/intel_gna/src/frontend/layer_quantizer.cpp
+++ b/src/plugins/intel_gna/src/frontend/layer_quantizer.cpp
@@ -90,7 +90,7 @@ size_t LayerQuantizer::GetBiasSizeForLayer(InferenceEngine::WeightableLayer& wl)
        return wl._biases->size();
    } else if (LayerInfo(wl).isConvolution()) {
        // Calculating biases len using outdata dims: biases number should be equal to output channels number
-        return InferenceEngine::GetDataDimByName(wl.outData.front(), InferenceEngine::DataDimName::C);
+        return InferenceEngine::GetDataDimSizeNHWC(wl.outData.front(), InferenceEngine::DataDimName::C);
    } else {
        // Calculating biases size using outData dimensions
        return wl.outData.front()->getDims().back();
--- a/src/plugins/intel_gna/src/frontend/scale_factor_calc.cpp
+++ b/src/plugins/intel_gna/src/frontend/scale_factor_calc.cpp
@@ -1265,7 +1265,7 @@ bool ScaleFactorCalculator::ScaleFactorPerLayerWeightable(InferenceEngine::Weigh
        double weights_reducer = 1.0;
        auto conv = dynamic_cast<InferenceEngine::ConvolutionLayer*>(wl);
        if (conv && !LayerInfo(conv).isConvolutionFilter()) {
-            const auto inDepth = GetDataDimByName(conv->insData.front().lock(), InferenceEngine::DataDimName::C);
+            const auto inDepth = GetDataDimSizeNHWC(conv->insData.front().lock(), InferenceEngine::DataDimName::C);
            weights_reducer = gna_convolution_layer::getWeightsReducer(*conv);
            weights_reducer *= MAX_VAL_2B_FEAT * scaleRange * inDepth / std::numeric_limits<int32_t>::max();
            weights_reducer = std::max(1.0, weights_reducer);
--- a/src/plugins/intel_gna/src/gna_graph_compiler.cpp
+++ b/src/plugins/intel_gna/src/gna_graph_compiler.cpp
@@ -307,12 +307,26 @@ void GNAGraphCompiler::ConstPrimitive(InferenceEngine::CNNLayerPtr constLayer) {

 void GNAGraphCompiler::assertConvolutionLayoutProper(const InferenceEngine::DataPtr& data) {
    if (data->getLayout() != InferenceEngine::Layout::NHWC && data->getLayout() != InferenceEngine::Layout::NCHW &&
-        data->getLayout() != InferenceEngine::Layout::NC) {
+        data->getLayout() != InferenceEngine::Layout::NC && data->getLayout() != InferenceEngine::Layout::CHW) {
        THROW_GNA_EXCEPTION << "layer: \"Convolution\" with layout " << data->getLayout()
                            << " isn't currently supported on GNA";
    }
 }

+namespace {
+
+template <typename T>
+PropertyVector<T> property_vector_append(PropertyVector<T> properties, T value) {
+    std::vector<T> new_values;
+    for (size_t i = 0; i < properties.size(); ++i)
+        new_values.push_back(properties[i]);
+    new_values.push_back(value);
+
+    return PropertyVector<T>(new_values);
+}
+
+}  // namespace
+
 /**
 * Create AMIntelDNN Convolutional1DComponent from ConvolutionLayer
 *
@@ -338,15 +352,24 @@ void GNAGraphCompiler::ConvolutionPrimitive(InferenceEngine::CNNLayerPtr layer)
    const auto outputs = layer->outData.front();
    assertConvolutionLayoutProper(inputs);

-    const auto in_batch = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::N);
-    const auto in_channels = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::C);
-    auto in_height = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::H);
-    auto in_width = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::W);
+    const auto in_batch = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::N);
+    const auto in_channels = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::C);
+    auto in_height = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::H);
+    auto in_width = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::W);
+    const auto out_batch = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::N);
+    const auto out_channels = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::C);
+    auto out_height = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::H);
+    auto out_width = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::W);

-    const auto out_batch = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::N);
-    const auto out_channels = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::C);
-    auto out_height = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::H);
-    auto out_width = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::W);
+    if (inputs->getLayout() == InferenceEngine::Layout::CHW) {
+        // convolution is ngraph-3D here. Make some fixes to work with it as it's ngraph-4D
+        convolution._kernel_y = 1;
+        convolution._dilation_y = 1;
+        convolution._stride_y = 1;
+
+        convolution._padding = property_vector_append<unsigned int>(convolution._padding, 0);
+        convolution._pads_end = property_vector_append<unsigned int>(convolution._pads_end, 0);
+    }

    if (in_height > 1 && in_width == 1 && !ShouldUseOnlyConv2DGnaIface()) {
        std::swap(in_height, in_width);
@@ -589,42 +612,12 @@ void GNAGraphCompiler::finalizeConvolution1DPrimitive(InferenceEngine::CNNLayerP
        });
    }

-    // TODO: convolution might be not the first layer in sorted order but connected via split for example - dont know
-    // how kaldi will handle that
-    if (!dnn->do_rotate_input) {
-        if ((inputs->getLayout() != InferenceEngine::Layout::NHWC || transpose_h_w) &&
-            LayerInfo(connectedInputLayer).isInput()) {
-            //  Kaldi features are opposite orientation
-            dnn->do_rotate_input = true;
-            dnn->num_rotate_rows = effectiveStride;
-            dnn->num_rotate_columns = num_inputs / effectiveStride;
-        } else {
-            dnn->do_rotate_input = false;
-        }
-    }
-
    connectOutput(layer, ptr_outputs, num_data_bytes_out);

-    // Transpose H with W or C with HW
-    auto A = transpose_h_w ? in_kernel_h : in_channels;
-    auto B = transpose_h_w ? in_kernel_w : convolution._kernel[X_AXIS];
-
-    std::vector<uint8_t> transposedWeights;
-    for (uint32_t k = 0; k < num_filters; k++) {
-        uint8_t* ptr_filt_current =
-            convolution._weights->cbuffer().as<uint8_t*>() + k * A * B * convolution.precision.size();
-        auto transposedPart = transposeMatrix(ptr_filt_current, convolution.precision.size(), A, B);
-        transposedWeights.insert(transposedWeights.end(), transposedPart.begin(), transposedPart.end());
-    }
-    if (transposedWeights.size() != convolution._weights->byteSize()) {
-        THROW_GNA_LAYER_EXCEPTION(&convolution) << "weights was transposed incorrectly. " << transposedWeights.size()
-                                                << ' ' << convolution._weights->byteSize();
-    }
-
    if (num_conv_kernel_padding == 0) {
        gnamem->getQueue(REGION_RO)->push_local_ptr(layer,
                                                    ptr_weights,
-                                                    transposedWeights.data(),
+                                                    convolution._weights->cbuffer(),
                                                    convolution._weights->byteSize());
    } else {
        auto paddedWeights = num_filter_coefficients * num_filters;
@@ -636,7 +629,7 @@ void GNAGraphCompiler::finalizeConvolution1DPrimitive(InferenceEngine::CNNLayerP
                            layerName,
                            num_conv_kernel_padding,
                            cpSize,
-                            transposedWeights,
+                            convolution,
                            num_filters,
                            single_conv_kernel_size](void* data, std::size_t size) {
            if (paddedWeightsSize > size) {
@@ -648,7 +641,7 @@ void GNAGraphCompiler::finalizeConvolution1DPrimitive(InferenceEngine::CNNLayerP
            for (uint32_t i = 0; i < num_filters; i++) {
                ie_memcpy(dstPtr + offset,
                          size - offset,
-                          transposedWeights.data() + single_conv_kernel_size * i * cpSize,
+                          convolution._weights->cbuffer().as<uint8_t*>() + single_conv_kernel_size * i * cpSize,
                          single_conv_kernel_size * cpSize);
                offset += single_conv_kernel_size * cpSize;
                ie_memcpy(dstPtr + offset, size - offset, &padding_zeros[0], padding_zeros.size());
@@ -783,22 +776,6 @@ void GNAGraphCompiler::finalizeConvolution2DPrimitive(InferenceEngine::CNNLayerP

    auto connectedInputLayer = connectInput(layer, ptr_inputs, num_data_bytes_in).input;

-    // TODO: convolution might be not the first layer in sorted order but connected via split for example - dont know
-    // how kaldi will handle that
-    if (!dnn->do_rotate_input && inputs->getLayout() != InferenceEngine::Layout::NHWC &&
-        LayerInfo(connectedInputLayer).isInput()) {
-        //  Kaldi features are opposite orientation
-        dnn->do_rotate_input = true;
-        dnn->num_rotate_rows = in_channels;
-        if (in_height != 1) {
-            dnn->num_rotate_rows *= convolution._stride_y;
-        }
-        if (in_width != 1) {
-            dnn->num_rotate_rows *= convolution._stride_x;
-        }
-        dnn->num_rotate_columns = num_inputs / dnn->num_rotate_rows;
-    }
-
    connectOutput(layer, ptr_outputs, num_data_bytes_out);

    const auto convolution_precision = convolution.precision.size();
@@ -815,7 +792,7 @@ void GNAGraphCompiler::finalizeConvolution2DPrimitive(InferenceEngine::CNNLayerP
        ALIGN(effective_single_kernel_size, Limitations::kConvEachKernelByteAlignment) - effective_single_kernel_size;
    for (uint32_t k = 0; k < convolution._out_depth; k++) {
        uint8_t* ptr_filt_current = convolution._weights->cbuffer().as<uint8_t*>() + k * single_kernel_size;
-        auto transposed_part = transposeMatrix(ptr_filt_current, convolution_precision, in_channels, kernelHW);
+        auto transposed_part = copy_matrix(ptr_filt_current, convolution.precision.size(), in_channels, kernelHW);
        transposed_weights.insert(transposed_weights.end(), transposed_part.begin(), transposed_part.end());
        transposed_weights.resize(transposed_weights.size() + effective_single_kernel_size - single_kernel_size +
                                  kernel_pad);
@@ -997,13 +974,19 @@ void GNAGraphCompiler::PoolingPrimitive(InferenceEngine::CNNLayerPtr layer) {
    auto inputs = layer->insData.begin()->lock();
    auto outputs = *layer->outData.begin();

-    uint32_t w_dim_in = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::W);
-    uint32_t h_dim_in = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::H);
-    const uint32_t c_dim_in = InferenceEngine::GetDataDimByName(inputs, InferenceEngine::DataDimName::C);
+    uint32_t w_dim_in = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::W);
+    uint32_t h_dim_in = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::H);
+    const uint32_t c_dim_in = GetDataDimSizeNHWC(inputs, InferenceEngine::DataDimName::C);

-    uint32_t w_dim_out = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::W);
-    uint32_t h_dim_out = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::H);
-    const uint32_t c_dim_out = InferenceEngine::GetDataDimByName(outputs, InferenceEngine::DataDimName::C);
+    uint32_t w_dim_out = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::W);
+    uint32_t h_dim_out = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::H);
+    const uint32_t c_dim_out = GetDataDimSizeNHWC(outputs, InferenceEngine::DataDimName::C);
+
+    if (inputs->getLayout() == InferenceEngine::Layout::CHW) {
+        // Pooling is ngraph-3D here. Make some fixes to work with it as it's ngraph-4D
+        pooling._kernel = property_vector_append<unsigned int>(pooling._kernel, 1);
+        pooling._stride = property_vector_append<unsigned int>(pooling._stride, 1);
+    }

    void* ptr_inputs = nullptr;
    void* ptr_outputs = nullptr;
@@ -2590,7 +2573,8 @@ void GNAGraphCompiler::connectOutput(InferenceEngine::CNNLayerPtr layer, void* p
                if (layer->params.find("output_offset") != layer->params.end()) {
                    output_offset = layer->GetParamAsInt("output_offset");
                }
-                gnamem->getQueue(REGION_AUTO)->bind_ptr(layer, ptr, &concatLayerInfoItem.gna_ptr, output_offset);
+                gnamem->getQueue(REGION_AUTO)
+                    ->bind_ptr(layer, ptr, &concatLayerInfoItem.gna_ptr, output_offset, num_data_bytes_out);
            }
            return;
        }
@@ -2859,5 +2843,15 @@ std::vector<uint8_t> GNAGraphCompiler::transposeMatrix(uint8_t* ptr_matrix,
    return temp_buffer;
 }

+std::vector<uint8_t> GNAGraphCompiler::copy_matrix(uint8_t* ptr_matrix,
+                                                   size_t element_size,
+                                                   uint32_t num_rows,
+                                                   uint32_t num_cols) {
+    const size_t dest_size = num_rows * num_cols * element_size;
+    std::vector<uint8_t> temp_buffer(dest_size);
+    ::memcpy(temp_buffer.data(), ptr_matrix, dest_size);
+    return temp_buffer;
+}
+
 }  // namespace intel_gna
 }  // namespace ov
--- a/src/plugins/intel_gna/src/gna_graph_compiler.hpp
+++ b/src/plugins/intel_gna/src/gna_graph_compiler.hpp
@@ -53,6 +53,10 @@ private:
                                                size_t element_size,
                                                uint32_t num_rows,
                                                uint32_t num_cols);
+    std::vector<uint8_t> static copy_matrix(uint8_t* ptr_matrix,
+                                            size_t element_size,
+                                            uint32_t num_rows,
+                                            uint32_t num_cols);

    bool ShouldUseOnlyConv2DGnaIface() const;

--- a/src/plugins/intel_gna/src/gna_graph_patterns.hpp
+++ b/src/plugins/intel_gna/src/gna_graph_patterns.hpp
@@ -237,32 +237,6 @@ inline InferenceEngine::CNNLayerPtr FindPermutationAfterConvolutionInKaldiModel(
    return next;
 }

-/**
- * @brief identifies if a model must be converted to NHWC, it must not be neither NHWC, nor Kaldi
- * @param layers model sorted layers
- */
-inline bool MustBeConvertedFromNCHWToNHWC(const std::vector<InferenceEngine::CNNLayerPtr>& layers) {
-    for (auto& l : layers) {
-        if (!LayerInfo(l).isConvolution())
-            continue;
-
-        InferenceEngine::CNNLayerPtr next;
-        std::tie(std::ignore, next) = FindPermutationsAroundConvolutionInNHWCModel(l);
-        if (next != nullptr)
-            return false;
-        // If a convolution has only 1-dimension input and output we should skip it
-        auto in_dims = l->insData.begin()->lock()->getDims();
-        auto out_dims = l->outData.front()->getDims();
-
-        if (ov::intel_gna::graph_utils::is_one_dim_shapes(in_dims, out_dims)) {
-            continue;
-        }
-
-        return FindPermutationAfterConvolutionInKaldiModel(l) == nullptr;
-    }
-    return false;
-}
-
 /**
 * @brief returns transposition information for a layer based on the previous convolution or pooling dimensions order
 * @param layer layer from which transposition info search must be started
--- a/src/plugins/intel_gna/src/gna_graph_tools.hpp
+++ b/src/plugins/intel_gna/src/gna_graph_tools.hpp
@@ -924,4 +924,38 @@ inline uint32_t GetDataDimByName(InferenceEngine::DataPtr data, DataDimName dimN
    return GetDimFromBack(dims, backOffsets[dimIxInNCHW]);
 }

+/**
+ * @brief returns a size of a specified data dimension depending on the layout
+ *        NHWC specialization
+ * @param data a pointer to the data
+ * @param dimName dimension name
+ */
+inline uint32_t GetDataDimSizeNHWC(InferenceEngine::DataPtr data, DataDimName dimName) {
+    uint32_t dimIxInNCHW = static_cast<uint32_t>(dimName);
+    IE_ASSERT(dimIxInNCHW <= 3);
+
+    std::vector<uint32_t> backOffsets;
+    switch (data->getLayout()) {
+    case Layout::C:
+    case Layout::NC:
+        // 1 will be returned for offsets > 2
+        backOffsets = std::vector<uint32_t>{2, 1, 3, 4};
+        break;
+    case Layout::HWC:
+        // 1 will be returned for offset 4
+    case Layout::NHWC:
+        backOffsets = std::vector<uint32_t>{4, 3, 2, 1};
+        break;
+    case Layout::CHW:
+        // 1 will be returned for offset 4
+    case Layout::NCHW:
+        backOffsets = std::vector<uint32_t>{4, 1, 3, 2};
+        break;
+    default:
+        THROW_GNA_EXCEPTION << data->getName() << " Unexpected layout " << data->getLayout();
+    }
+    auto dims = data->getDims();
+    return GetDimFromBack(dims, backOffsets[dimIxInNCHW]);
+}
+
 }  // namespace InferenceEngine
--- a/src/plugins/intel_gna/src/gna_plugin.cpp
+++ b/src/plugins/intel_gna/src/gna_plugin.cpp
@@ -344,13 +344,13 @@ void GNAPlugin::PrePostProcess(InferenceEngine::Blob::Ptr input_blob,
                               std::shared_ptr<ov::Model> model) {
    const ov::element::Type input_type = details::convertPrecision(input_blob->getTensorDesc().getPrecision());
    const ov::element::Type output_type = details::convertPrecision(output_blob->getTensorDesc().getPrecision());
-    const ov::Shape& input_shape = model->get_parameters().front()->get_shape();
-    const ov::Shape& output_shape = model->get_results().front()->get_shape();
+    const ov::Shape& output_shape = output_blob->getTensorDesc().getDims();

    for (const auto& param : model->get_parameters()) {
        param->set_element_type(input_type);
    }
    model->validate_nodes_and_infer_types();
+    const ov::Shape& input_shape = model->get_parameters()[0]->get_output_shape(0);

    ov::TensorVector inputs = {ov::Tensor(input_type, input_shape, input_blob->cbuffer().as<void*>())};
    ov::TensorVector results = {ov::Tensor(output_type, output_shape, output_blob->buffer().as<void*>())};
@@ -611,52 +611,6 @@ bool GNAPlugin::TryToInitOutput(const std::string& portName, InferenceEngine::CN
    return false;
 }

-void GNAPlugin::FillInputsAndOutputsTranspositionInfo(const InferenceEngine::CNNNetwork& net) {
-    OV_ITT_SCOPED_TASK(itt::domains::GNA_LT, "FillInputsAndOutputsTranspositionInfo");
-    auto printTranspositionInfo = [](const std::vector<TranspositionInfo>& transpositionInfo) {
-        for (const auto& transpositionInfoPart : transpositionInfo) {
-            log::debug() << "transpose=" << transpositionInfoPart.transpose
-                         << " rows_num=" << transpositionInfoPart.num_transpose_rows
-                         << " columns_num=" << transpositionInfoPart.num_transpose_columns << "\n";
-        }
-    };
-
-    auto inputLayers = CNNNetGetAllInputLayers(net);
-    for (const auto& inputLayer : inputLayers) {
-        // Collect information for inputs transposition
-        if (!LayerInfo(inputLayer).isInput())
-            continue;
-        auto transpositionInfo = FindTranspositionInfoFromNextLayers(inputLayer);
-        if (transpositionInfo.empty())
-            continue;
-
-        transpose_inputs_info.insert({inputLayer->name, transpositionInfo});
-        log::debug() << "Input " << inputLayer->name << " transposition info: \n";
-        printTranspositionInfo(transpositionInfo);
-    }
-
-    auto outputsMap = net.getOutputsInfo();
-    for (const auto& outPort : outputsMap) {
-        auto outLayer = getCreatorLayer(outPort.second).lock();
-        // Collect information for outputs transposition
-        if (!LayerInfo(outLayer).isOutput())
-            continue;
-        auto transpositionInfo = FindTranspositionInfoFromPrevLayers(outLayer);
-        if (transpositionInfo.empty())
-            continue;
-
-        // Swap transposition info rows and columns since we need to transpose output back from NHWC to NCHW
-        for (auto&& transpositionInfoPart : transpositionInfo) {
-            if (transpositionInfoPart.transpose) {
-                std::swap(transpositionInfoPart.num_transpose_rows, transpositionInfoPart.num_transpose_columns);
-            }
-        }
-        transpose_outputs_info.insert({outLayer->name, transpositionInfo});
-        log::debug() << "Output " << outLayer->name << " transposition info: \n";
-        printTranspositionInfo(transpositionInfo);
-    }
-}
-
 #ifdef PLOT
 void GNAPlugin::AddDebugProperties(const InferenceEngine::CNNLayerPtr layer,
                                   InferenceEngine::ordered_properties& printed_properties,
@@ -751,10 +705,6 @@ void GNAPlugin::LoadNetwork(const CNNNetwork& _network) {
        UpdateInputScaleFromNetwork(network);
    }

-    if (MustBeConvertedFromNCHWToNHWC(CNNNetSortTopologically(network))) {
-        FillInputsAndOutputsTranspositionInfo(network);
-    }
-
    InferenceEngine::CNNNetwork newNet;

    if (gnaFlags->sw_fp32) {
@@ -995,22 +945,6 @@ void GNAPlugin::LoadNetwork(const CNNNetwork& _network) {
        }
    }

-    if (dnn->do_rotate_input && transpose_inputs_info.empty()) {
-        for (auto& inputLayer : inputLayers) {
-            transpose_inputs_info.insert(
-                {inputLayer->name,
-                 {TranspositionInfo{dnn->do_rotate_input, dnn->num_rotate_rows, dnn->num_rotate_columns}}});
-        }
-    }
-
-    // TODO: Need to remove this conversion when ngraph NCHW->NHWC transformation is enabled
-    if (!transpose_inputs_info.empty()) {
-        ConvertTransposeMapToModel(transpose_inputs_info, inputs_ptr_->Get());
-    }
-    if (!transpose_outputs_info.empty()) {
-        ConvertTransposeMapToModel(transpose_outputs_info, outputs_.Get());
-    }
-
    DumpXNNToFile();

 #ifdef PLOT
--- a/src/plugins/intel_gna/src/gna_plugin.hpp
+++ b/src/plugins/intel_gna/src/gna_plugin.hpp
@@ -37,6 +37,8 @@ class WorkerPool;
 class Worker;
 }  // namespace request

+using namespace ov::intel_gna::pre_post_processing;
+
 class GNAPlugin : public InferenceEngine::IInferencePlugin {
 protected:
    std::string _pluginName = "GNA";
@@ -204,6 +206,13 @@ protected:
                        InferenceEngine::Blob::Ptr output_blob,
                        std::shared_ptr<ov::Model> model);

+    /**
+     * Run ngraph model on CPU to modify inputs/outputs
+     */
+    void pre_post_process(InferenceEngine::Blob::Ptr input_blob,
+                          InferenceEngine::Blob::Ptr output_blob,
+                          std::shared_ptr<ov::Model> model);
+
    void ImportFrames(void* ptr_dst,
                      const void* ptr_src,
                      InferenceEngine::Precision input_precision,
@@ -246,14 +255,6 @@ protected:
     * @return true if the output is initiated, false otherwise
     */
    bool TryToInitOutput(const std::string& portName, InferenceEngine::CNNLayerPtr layer);
-
-    /**
-     * @brief Fills inputs and outputs transposition info for model convertion from NCHW to NHWC.
-     *        Information for transposition is found from convolution/pooling input or output dimensions.
-     * @param layers model sorted layers
-     */
-    void FillInputsAndOutputsTranspositionInfo(const InferenceEngine::CNNNetwork& net);
-
    bool isFP32ModeActive() const;
    std::shared_ptr<request::ModelWrapper> createModelWrapperForLoadNetwork(bool trivial);
    std::shared_ptr<request::ModelWrapper> createModelWrapperForImportNetwork(uint32_t numberOfOperations);
--- a/src/plugins/intel_gna/src/gna_transformations_pipeline.cpp
+++ b/src/plugins/intel_gna/src/gna_transformations_pipeline.cpp
@@ -7,6 +7,8 @@
 #include "gna_itt.hpp"
 #include "legacy/net_pass.h"
 #include "legacy/transformations/convert_opset1_to_legacy/convert_opset1_to_legacy.hpp"
+#include "legacy/transformations/convert_opset1_to_legacy/convert_strided_slice_to_crop.hpp"
+#include "ngraph/opsets/opset2.hpp"
 #include "ngraph/opsets/opset7.hpp"
 #include "openvino/pass/manager.hpp"
 #include "optimizer/gna_pass_manager.hpp"
@@ -18,7 +20,9 @@
 #include "transformations/common_optimizations/fq_reshape_fusion.hpp"
 #include "transformations/common_optimizations/pull_transpose_through_fq.hpp"
 #include "transformations/common_optimizations/relu_fake_quantize_fusion.hpp"
+#include "transformations/common_optimizations/reshape_sequence_fusion.hpp"
 #include "transformations/common_optimizations/transpose_sinking.hpp"
+#include "transformations/common_optimizations/transpose_to_reshape.hpp"
 #include "transformations/control_flow/unroll_tensor_iterator.hpp"
 #include "transformations/convert_dwsc_to_scaleshifts.hpp"
 #include "transformations/convert_matmul_to_pointwise_convolution.hpp"
@@ -28,6 +32,8 @@
 #include "transformations/decompose_mvn.hpp"
 #include "transformations/fp16_compression/convert_compression_only_to_legacy.hpp"
 #include "transformations/fp16_compression/mark_decompression_convert_constant_folding.hpp"
+#include "transformations/fuse_conv_bias_activation.hpp"
+#include "transformations/gather_sinking.hpp"
 #include "transformations/handle_transposes_around_matmul.hpp"
 #include "transformations/init_node_info.hpp"
 #include "transformations/insert_copy_layer.hpp"
@@ -37,6 +43,7 @@
 #include "transformations/markup_fusable_transpose.hpp"
 #include "transformations/op_conversions/convert_mvn1_to_mvn6.hpp"
 #include "transformations/op_conversions/convert_sequences_to_tensor_iterator.hpp"
+#include "transformations/op_conversions/convert_slice_to_strided_slice.hpp"
 #include "transformations/op_conversions/gru_cell_decomposition.hpp"
 #include "transformations/op_conversions/lstm_cell_decomposition.hpp"
 #include "transformations/op_conversions/softsign_decomposition.hpp"
@@ -48,13 +55,28 @@
 #include "transformations/remove_in_out_processing.hpp"
 #include "transformations/remove_single_input_concat.hpp"
 #include "transformations/reorder_activation_and_pooling.hpp"
+#include "transformations/replace_gna_nhwc_layers.hpp"
+#include "transformations/reshape_transpose_substitute.hpp"
+#include "transformations/rotate_inputs.hpp"
 #include "transformations/split_convolution_with_large_buffer_size.hpp"
 #include "transformations/split_eltwise.hpp"
 #include "transformations/substitute_softsign.hpp"
 #include "transformations/swap_input_matmul_gna.hpp"
+#include "transformations/transpose_sinking/ts_concat.hpp"
+#include "transformations/transpose_sinking/ts_fuse.hpp"
+#include "transformations/transpose_sinking/ts_general.hpp"
+#include "transformations/transpose_sinking/ts_split.hpp"
+#include "transformations/ts_concat_forward.hpp"
+#include "transformations/ts_split_backward.hpp"
 #include "transformations/unfuse_reshape_and_transpose.hpp"
+#include "transformations/utils/transformation_helper.hpp"
 #include "transformations/utils/utils.hpp"

+using namespace ov;
+using namespace ov::opset8;
+using namespace ov::intel_gna::limitations;
+using namespace ov::intel_gna::pass::helper;
+
 namespace ov {
 namespace intel_gna {

@@ -64,12 +86,13 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,

    fake_quantized = ov::op::util::has_op_with_type<ngraph::op::FakeQuantize>(model);
    const bool has_convolution = ov::op::util::has_op_with_type<ngraph::opset7::Convolution>(model);
+    const bool has_maxpool = ov::op::util::has_op_with_type<ov::opset8::MaxPool>(model);
+    const bool has_slice = ov::op::util::has_op_with_type<ov::opset8::Slice>(model);
    const bool has_matmul = ov::op::util::has_op_with_type<ngraph::opset7::MatMul>(model);
-    const bool has_mvn = ov::op::util::has_op_with_type<ngraph::opset7::MVN>(model) ||
+    const bool has_mvn = ov::op::util::has_op_with_type<ov::opset8::MVN>(model) ||
                         ov::op::util::has_op_with_type<ov::op::v0::MVN>(model);
    ov::pass::Manager manager;
    manager.register_pass<ov::pass::InitNodeInfo>();
-
    // In OV API 2.0(IRv10) default convertion to fp32 (inputs, outputs and weights) is disabled
    // and we need to run the ConvertPrecision transformation to support old networks.
    manager.register_pass<ov::pass::ConvertPrecision>(precisions_map{{ngraph::element::f16, ngraph::element::f32}});
@@ -104,7 +127,6 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
    manager.register_pass<ov::intel_gna::pass::SwapInputMatMulWithBias>();
    manager.register_pass<ov::intel_gna::pass::SwapInputMatMul>();
    manager.register_pass<ov::intel_gna::pass::HandleTransposesAroundMatMul>();
-    manager.register_pass<ov::intel_gna::pass::InsertTransposeAfterConvOrPool>();
    manager.register_pass<ov::intel_gna::pass::Unfuse2dto4dReshapeAndTranspose>();
    manager.register_pass<ov::intel_gna::pass::Unfuse4dto2dReshapeAndTranspose>();
    manager.register_pass<ov::intel_gna::pass::RemoveExtraReshapes>();
@@ -112,11 +134,21 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
    manager.register_pass<ov::intel_gna::pass::RemoveSingleInputConcat>();
    manager.register_pass<ov::intel_gna::pass::SubstituteSoftsign>();
    manager.register_pass<ov::intel_gna::pass::InsertCopyBeforeLayerToBeEliminated>();
-    if (!has_convolution && !has_matmul && !has_mvn) {
-        // TODO: Remove this condition when the legacy layout transformation (NCHW->NHWC) is disabled
-        manager.register_pass<ov::intel_gna::pass::RemoveInputsProcessing>(input_output_subgraphs);
-        manager.register_pass<ov::intel_gna::pass::RemoveOutputsProcessing>(input_output_subgraphs);
+    // TODO enable this transformation for networks without convolutions
+    if (has_convolution || has_maxpool || has_mvn || has_matmul) {
+        manager.register_pass<ov::intel_gna::pass::ReplaceGnaNHWCLayers>();
+        manager.register_pass<ov::intel_gna::pass::InsertConvolutionTransposeHW>();
+        manager.register_pass<ov::pass::TransposeSinkingGeneral>();
+        manager.register_pass<ov::intel_gna::pass::GatherSinkingGeneral>();
+        manager.register_pass<ov::pass::ReshapeSequenceFusion>();
+        manager.register_pass<ov::pass::TransposeToReshape>();
+        manager.register_pass<ov::intel_gna::pass::GnaConvolutionFusion>();
+        manager.register_pass<ov::intel_gna::pass::TSConcatForward>();
+        manager.register_pass<ov::intel_gna::pass::TSSplitBackward>();
+        manager.register_pass<ov::pass::transpose_sinking::TSFuse>();
    }
+    manager.register_pass<ov::intel_gna::pass::RemoveInputsProcessing>(input_output_subgraphs);
+    manager.register_pass<ov::intel_gna::pass::RemoveOutputsProcessing>(input_output_subgraphs);
    manager.register_pass<ov::pass::ConvertOpSet3ToOpSet2>();
    manager.register_pass<ov::pass::ConvertOpSet2ToOpSet1>();
    manager.register_pass<ngraph::pass::ConvertOpSet1ToLegacy>();
@@ -160,6 +192,62 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
                                                                     {ov::element::u32, ov::element::i32}});
    const auto& pass_config = manager.get_pass_config();

+    pass_config->set_callback<ov::pass::transpose_sinking::TSConcatForward>(
+        [](const std::shared_ptr<const ov::Node>& node) -> bool {
+            const TransposeInfo transpose_info = get_first_input_transpose(node);
+            if (transpose_info.isEmpty())
+                return false;
+            const bool is_supported = Limitations::is_forward_transposed_concat_supported(
+                node,
+                transpose_info.transpose_const->get_axis_vector_val());
+            if (!is_supported)
+                mark_input_transposes_as_nosinking(node);
+            return !is_supported;
+        });
+
+    pass_config->set_callback<ov::pass::transpose_sinking::TSConcatBackward>(
+        [](const std::shared_ptr<const ov::Node>& node) -> bool {
+            const TransposeInfo transpose_info = get_first_output_transpose(node);
+            if (transpose_info.isEmpty())
+                return false;
+            return !Limitations::is_backward_transposed_concat_supported(
+                node,
+                transpose_info.transpose_const->get_axis_vector_val());
+        });
+
+    pass_config->set_callback<ov::pass::transpose_sinking::TSSplitForward>(
+        [](const std::shared_ptr<const ov::Node>& node) -> bool {
+            const TransposeInfo transpose_info = get_first_input_transpose(node);
+            if (transpose_info.isEmpty())
+                return false;
+            const bool is_supported = Limitations::is_forward_transposed_split_supported(
+                node,
+                transpose_info.transpose_const->get_axis_vector_val());
+            if (!is_supported)
+                mark_input_transposes_as_nosinking(node);
+            return !is_supported;
+        });
+
+    pass_config->set_callback<ov::pass::transpose_sinking::TSSplitBackward>(
+        [](const std::shared_ptr<const ov::Node>& node) -> bool {
+            const TransposeInfo transpose_info = get_first_output_transpose(node);
+            if (transpose_info.isEmpty())
+                return false;
+            return !Limitations::is_backward_transposed_split_supported(
+                node,
+                transpose_info.transpose_const->get_axis_vector_val());
+        });
+
+    /**
+     * TransposeSinking doesn't currently support StridedSlice. We disable SliceToStridedSlice
+     * transformation to prevent convert Slice to StridedSlice. This allows us to work with
+     * networks, that initialy have Slice.
+     * That could be removed after StridedSlice implementation in TransposeSinking
+     */
+    if (has_slice && (has_convolution || has_maxpool || has_mvn)) {
+        pass_config->disable<ov::pass::SliceToStridedSlice>();
+    }
+
    // Allowing FP16 Converts to be folded and FP16 constants to upgrade to FP32 data type
    pass_config->disable<ov::pass::ConvertCompressedOnlyToLegacy>();
    pass_config->disable<ov::pass::DisableDecompressionConvertConstantFolding>();
@@ -177,8 +265,23 @@ void TransformationsPipeline::apply(const std::shared_ptr<ov::Model>& model,
    // Operations Max and Min aren't supported
    pass_config->disable<ov::pass::ConcatReduceFusion>();

+    pass_config->disable<ov::pass::ConcatReduceFusion>();
    manager.run_passes(model);

+    /**
+     * As we disabled SliceToStridedSlice, we have after all transformations
+     * Slice, that is not supported natively in our plugin. Here we convert
+     * Slice -> StridedSlice -> CropIE
+     * That could be removed after StridedSlice implementation in TransposeSinking
+     */
+    if (has_slice && (has_convolution || has_maxpool || has_mvn)) {
+        ov::pass::Manager manager;
+        manager.register_pass<ov::pass::InitNodeInfo>();
+        manager.register_pass<ov::pass::SliceToStridedSlice>(true);
+        manager.register_pass<ngraph::pass::ConvertStridedSliceToCropMatcher>();
+        manager.run_passes(model);
+    }
+
    is_ngraph_passes_used = true;
 }

@@ -204,8 +307,6 @@ void TransformationsPipeline::apply_legacy(const InferenceEngine::CNNNetwork& ne
    passes->registerPass<FuseFQIntoWeightsPass>();
    passes->registerPass<MoveFakeQuantizeLayerIntoQuantParamsPass>();

-    passes->registerPass<TransposeWeightsFromNCHWToNHWCPass>();
-
    passes->registerPass<SubstitutePReluPass>();

    if (!is_ngraph_passes_used) {
@@ -221,7 +322,7 @@ void TransformationsPipeline::apply_legacy(const InferenceEngine::CNNNetwork& ne
    passes->registerPass<FlattenTrivialConcatPass>();
    passes->registerPass<InsertConcatAligningFilterPass>();
    passes->registerPass<ReorderConcatInputsPass>();
-    passes->registerPass<RemovePermutationsNHWCToNCHWPass>();
+
    // Keep legacy inserting of Identity layer here
    // because concat and split aliging passes are not moved to ngraph yet
    passes->registerPass<InsertIdentityLayerPass>();
--- a/src/plugins/intel_gna/src/layers/gna_convolution_layer.cpp
+++ b/src/plugins/intel_gna/src/layers/gna_convolution_layer.cpp
@@ -59,11 +59,11 @@ double getWeightsReducer(InferenceEngine::ConvolutionLayer& conv) {
    const std::vector<KRT> reducers{{49, 3.0}, {36, 2.6}, {21, 2.3}, {14, 1.7}, {9, 1.3}, {7, 1.2}};
    auto reducer = 1.0;
    const auto inDepth =
-        InferenceEngine::GetDataDimByName(conv.insData.front().lock(), InferenceEngine::DataDimName::C);
+        InferenceEngine::GetDataDimSizeNHWC(conv.insData.front().lock(), InferenceEngine::DataDimName::C);
    const auto inHeight =
-        InferenceEngine::GetDataDimByName(conv.insData.front().lock(), InferenceEngine::DataDimName::H);
+        InferenceEngine::GetDataDimSizeNHWC(conv.insData.front().lock(), InferenceEngine::DataDimName::H);
    const auto inWidth =
-        InferenceEngine::GetDataDimByName(conv.insData.front().lock(), InferenceEngine::DataDimName::W);
+        InferenceEngine::GetDataDimSizeNHWC(conv.insData.front().lock(), InferenceEngine::DataDimName::W);
    if (is3DInputOr2DKernel(inHeight, inWidth, inDepth, conv._kernel_y, conv._kernel_x) &&
        !isMappableFrom2DTo1D(inHeight,
                              inWidth,
--- a/src/plugins/intel_gna/src/layers/gna_layer_info.hpp
+++ b/src/plugins/intel_gna/src/layers/gna_layer_info.hpp
@@ -297,7 +297,7 @@ public:
        return isOfType("FakeQuantize");
    }
    bool isNonFunctional() const {
-        return isOfType("reshape") || isOfType("squeeze") || isOfType("unsqueeze") || isTrivialPermute();
+        return isOfType("reshape") || isOfType("squeeze") || isOfType("unsqueeze") || isTrivialPermute() || is_gather();
    }
    bool isReshape() const noexcept {
        return isOfType("reshape");
@@ -305,6 +305,9 @@ public:
    bool isPermute() const noexcept {
        return isOfType("permute");
    }
+    bool is_gather() const noexcept {
+        return isOfType("gather");
+    }
    bool isPermuteFusable() const noexcept {
        return isPermute() &&
               (layer->params.count(ov::intel_gna::rt_info::GNATransposeFusable::get_type_info_static()) > 0);
@@ -349,11 +352,8 @@ public:
    bool isNonValuesChangable() const {
        return isNonFunctional() || isSplit() || isSlice() || isConcat();
    }
-    bool is_gather() const noexcept {
-        return isOfType("gather");
-    }
    bool is_fq_non_sensitive() const {
-        return isPermute() || is_gather() || isNonFunctional();
+        return isPermute() || isNonFunctional();
    }
    bool isPooling() const noexcept {
        return isOfType("pooling");
--- a/src/plugins/intel_gna/src/ops/gna_convolution.hpp
+++ b/src/plugins/intel_gna/src/ops/gna_convolution.hpp
@@ -133,6 +133,9 @@ public:
    void set_auto_pad(const ov::op::PadType& auto_pad) {
        m_auto_pad = auto_pad;
    }
+    bool has_add_node() const {
+        return m_has_add_node;
+    }
    bool has_bias() const {
        return m_has_add_node;
    }
--- a/src/plugins/intel_gna/src/optimizer/gna_pass_manager.cpp
+++ b/src/plugins/intel_gna/src/optimizer/gna_pass_manager.cpp
@@ -110,17 +110,6 @@ static void SumBlobs(Blob::Ptr& src_blob, Blob::Ptr& dst_blob) {
    }
 }

-static Blob::Ptr convertToRWBlob(const Blob::Ptr& readOnlyBlob, const std::string& name = {}) {
-    auto blob = Blob::CreateFromData(std::make_shared<Data>(name, readOnlyBlob->getTensorDesc()));
-    blob->allocate();
-    const auto ret = ie_memcpy(blob->buffer().as<uint8_t*>(),
-                               blob->size() * blob->getTensorDesc().getPrecision().size(),
-                               readOnlyBlob->buffer().as<uint8_t*>(),
-                               readOnlyBlob->size() * readOnlyBlob->getTensorDesc().getPrecision().size());
-    IE_ASSERT(ret == 0);
-    return blob;
-}
-
 // indexes stored in pass manager
 static const char identityLayersCounterName[] = "identityLayerCounter";
 static const char diagonalLayersCounterName[] = "diagonalLayerCounter";
@@ -2419,225 +2408,6 @@ void MoveFakeQuantizeLayerIntoQuantParamsPass ::run() {
    }
 }

-void TransposeWeightsFromNCHWToNHWCPass::run() {
-    OV_ITT_SCOPED_TASK(itt::domains::GNA_LT, "TransposeWeightsFromNCHWToNHWCPass");
-    if (!MustBeConvertedFromNCHWToNHWC(*pLayers))
-        return;
-
-    auto printTranspositionInfo = [](const std::vector<TranspositionInfo>& transpositionInfo) {
-        for (const auto& transpositionInfoPart : transpositionInfo) {
-            log::debug() << "transpose=" << transpositionInfoPart.transpose
-                         << " rows_num=" << transpositionInfoPart.num_transpose_rows
-                         << " columns_num=" << transpositionInfoPart.num_transpose_columns << "\n";
-        }
-    };
-
-    auto transpInfoMatchWeightsSize =
-        [](const std::vector<TranspositionInfo>& transpositionInfo, size_t weightsSize, const std::string& layerName) {
-            size_t totalElements = 0;
-            for (auto&& transpositionInfoPart : transpositionInfo) {
-                totalElements += transpositionInfoPart.num_transpose_rows * transpositionInfoPart.num_transpose_columns;
-            }
-            if (totalElements != weightsSize) {
-                THROW_GNA_EXCEPTION << layerName << " weights elements from transposition info (" << totalElements
-                                    << ") don't match input dimensions (" << weightsSize << ")";
-            }
-        };
-
-    for (auto&& l : *pLayers) {
-        if (LayerInfo(l).isScaleShift()) {
-            std::vector<TranspositionInfo> transpositionInfo;
-            // Try to find a convolution in previous layers
-            if (InferenceEngine::CNNNetHasPrevLayer(l.get())) {
-                transpositionInfo = FindTranspositionInfoFromPrevLayers(InferenceEngine::CNNNetPrevLayer(l));
-                // If no convolutions are found try to find them in next layers
-                if (!FoundPartToTranspose(transpositionInfo) && !l->outData.empty() &&
-                    !getInputTo(l->outData[0]).empty()) {
-                    transpositionInfo = FindTranspositionInfoFromNextLayers(getInputTo(l->outData[0]).begin()->second);
-                }
-            }
-            if (FoundPartToTranspose(transpositionInfo)) {
-                if (l->input()->getDims().front() > 1) {
-                    THROW_GNA_EXCEPTION << l->name
-                                        << " Weights transposition is not supported for a layer with batch size > 1";
-                }
-                auto weightable = dynamic_cast<WeightableLayer*>(l.get());
-                IE_ASSERT(weightable != nullptr);
-
-                size_t totalWeights = weightable->_weights->size();
-                transpInfoMatchWeightsSize(transpositionInfo, totalWeights, l->name);
-
-                ConvertTensorFromNCHWToNHWC(weightable->precision.size(),
-                                            1,
-                                            weightable->_weights->size(),
-                                            weightable->_weights->cbuffer().as<uint8_t*>(),
-                                            true,
-                                            transpositionInfo);
-                if (weightable->_biases) {
-                    ConvertTensorFromNCHWToNHWC(weightable->precision.size(),
-                                                1,
-                                                weightable->_biases->size(),
-                                                weightable->_biases->cbuffer().as<uint8_t*>(),
-                                                true,
-                                                transpositionInfo);
-                }
-                log::debug() << l->name << " weights and biases rows transposition info:\n";
-                printTranspositionInfo(transpositionInfo);
-            }
-        }
-
-        if (LayerInfo(l).isFullyConnected()) {
-            auto weightable = dynamic_cast<WeightableLayer*>(l.get());
-            IE_ASSERT(weightable != nullptr);
-            IE_ASSERT(weightable->_weights != nullptr);
-            auto precision = weightable->precision.size();
-            auto out_dims = l->outData[0]->getDims();
-            auto in_dims = l->input()->getDims();
-            auto weightsRows = InferenceEngine::details::product(std::begin(out_dims) + 1, std::end(out_dims));
-            auto weightsColumns = InferenceEngine::details::product(std::begin(in_dims) + 1, std::end(in_dims));
-            // Find a convolution in previous layers to rotate weights rows
-            if (InferenceEngine::CNNNetHasPrevLayer(l.get())) {
-                std::vector<TranspositionInfo> transpositionInfo;
-                auto prevLayer = InferenceEngine::CNNNetPrevLayer(l);
-                transpositionInfo = FindTranspositionInfoFromPrevLayers(prevLayer);
-                if (FoundPartToTranspose(transpositionInfo)) {
-                    if (l->input()->getDims().front() > 1) {
-                        THROW_GNA_EXCEPTION
-                            << l->name << " Weights transposition is not supported for a layer with batch size > 1";
-                    }
-                    if (LayerInfo(prevLayer).isSplit()) {
-                        // If we found a split it's not possible to rotate data
-                        THROW_GNA_EXCEPTION << l->name << " won't be transposed due to a split before it";
-                    }
-
-                    transpInfoMatchWeightsSize(transpositionInfo, weightsColumns, l->name);
-
-                    weightable->_weights = convertToRWBlob(weightable->_weights);
-
-                    ConvertTensorFromNCHWToNHWC(precision,
-                                                weightsRows,
-                                                weightsColumns,
-                                                weightable->_weights->buffer().as<uint8_t*>(),
-                                                true,
-                                                transpositionInfo);
-                    log::debug() << l->name << " weights rows transposition info:\n";
-                    printTranspositionInfo(transpositionInfo);
-                }
-            }
-            // Find a convolution in next layers to rotate weights columns
-            if (!l->outData.empty() && !getInputTo(l->outData[0]).empty()) {
-                std::vector<TranspositionInfo> transpositionInfo;
-                auto nextLayer = getInputTo(l->outData[0]).begin()->second;
-                transpositionInfo = FindTranspositionInfoFromNextLayers(nextLayer);
-                if (FoundPartToTranspose(transpositionInfo)) {
-                    if (l->outData[0]->getDims().front() > 1) {
-                        THROW_GNA_EXCEPTION
-                            << l->name << " Weights transposition is not supported for a layer with batch size > 1";
-                    }
-                    if (LayerInfo(nextLayer).isConcat()) {
-                        // If we found a concat it's not possible to rotate data
-                        THROW_GNA_EXCEPTION << l->name << " won't be transposed due to a concat after it";
-                    }
-
-                    transpInfoMatchWeightsSize(transpositionInfo, weightsRows, l->name);
-
-                    weightable->_weights = convertToRWBlob(weightable->_weights);
-
-                    ConvertTensorFromNCHWToNHWC(precision,
-                                                weightsRows,
-                                                weightsColumns,
-                                                weightable->_weights->cbuffer().as<uint8_t*>(),
-                                                false,
-                                                transpositionInfo);
-                    log::debug() << l->name << " weights columns transposition info:\n";
-                    printTranspositionInfo(transpositionInfo);
-                }
-            }
-        }
-
-        if (LayerInfo(l).isEltwise()) {
-            // We need to transpose a constant which is an eltwise input
-            auto firstInput = InferenceEngine::CNNNetPrevLayer(l, 0);
-            auto secondInput = InferenceEngine::CNNNetPrevLayer(l, 1);
-            if (!LayerInfo(firstInput).isConst() && !LayerInfo(secondInput).isConst()) {
-                continue;
-            }
-            // Let a constant to be the second input
-            if (LayerInfo(firstInput).isConst()) {
-                std::swap(firstInput, secondInput);
-            }
-            // Find a convolution in previous or next layers
-            auto transpositionInfo = FindTranspositionInfoFromPrevLayers(firstInput);
-            if (!FoundPartToTranspose(transpositionInfo) && !l->outData.empty() && !getInputTo(l->outData[0]).empty()) {
-                transpositionInfo = FindTranspositionInfoFromNextLayers(getInputTo(l->outData[0]).begin()->second);
-            }
-            if (FoundPartToTranspose(transpositionInfo)) {
-                auto blob = secondInput->blobs["custom"];
-                ConvertTensorFromNCHWToNHWC(blob->getTensorDesc().getPrecision().size(),
-                                            1,
-                                            blob->size(),
-                                            blob->buffer().as<uint8_t*>(),
-                                            true,
-                                            transpositionInfo);
-                log::debug() << secondInput->name << " data transposition info:\n";
-                printTranspositionInfo(transpositionInfo);
-            }
-        }
-
-        if (LayerInfo(l).isConcat()) {
-            auto concatLayer = LayerInfo(l).as<InferenceEngine::ConcatLayer*>();
-            IE_ASSERT(concatLayer != nullptr);
-            // If concatenation is along channel axis constant input transposition isn't required
-            if (concatLayer->_axis <= 1)
-                continue;
-
-            std::vector<InferenceEngine::CNNLayerPtr> constInputs;
-            bool transpose = false;
-            int nonConstInputIx = 0;
-            // Check if non-const inputs are transposed
-            for (int i = 0; InferenceEngine::CNNNetHasPrevLayer(l.get(), i); ++i) {
-                auto input = InferenceEngine::CNNNetPrevLayer(l, i);
-                if (LayerInfo(input).isConst()) {
-                    constInputs.push_back(input);
-                    continue;
-                }
-                auto transpositionInfo = FindTranspositionInfoFromPrevLayers(input);
-                bool transposeInput = FoundPartToTranspose(transpositionInfo);
-                if (nonConstInputIx == 0) {
-                    transpose = transposeInput;
-                } else if (transposeInput != transpose) {
-                    THROW_GNA_EXCEPTION << "Concat layer " << l->name << " inputs have different layouts";
-                }
-                ++nonConstInputIx;
-            }
-            if (!transpose)
-                continue;
-
-            // Transpose all constant inputs
-            for (auto&& input : constInputs) {
-                auto rows = GetDataDimByName(input->outData[0], DataDimName::C);
-                auto columns = GetDataDimByName(input->outData[0], DataDimName::H) *
-                               GetDataDimByName(input->outData[0], DataDimName::W);
-
-                auto blob = convertToRWBlob(input->blobs["custom"]);
-                input->blobs["custom"] = blob;
-
-                // A constant should have the same number of channels since concatenation will be in height/weight
-                // dimension
-                TranspositionInfo concatTranspositionInfo{true, rows, columns};
-                ConvertTensorFromNCHWToNHWC(blob->getTensorDesc().getPrecision().size(),
-                                            1,
-                                            blob->size(),
-                                            blob->buffer().as<uint8_t*>(),
-                                            true,
-                                            {concatTranspositionInfo});
-                log::debug() << input->name << " data transposition info:\n";
-                printTranspositionInfo({concatTranspositionInfo});
-            }
-        }
-    }
-}
-
 void FuseFullyConnectedWithEltwisePass::run() {
    // This legacy pass removes the Eltwise (only if it performs SUM op) from between FC and Any.
    // The blob data of Const layer attached to Eltwise is added to biases blob data of FC layer.
--- a/src/plugins/intel_gna/src/optimizer/gna_pass_manager.hpp
+++ b/src/plugins/intel_gna/src/optimizer/gna_pass_manager.hpp
@@ -214,14 +214,6 @@ DECL_PASS(FuseFQIntoWeights);
 */
 DECL_PASS(MoveFakeQuantizeLayerIntoQuantParams);

-/**
- * @brief convert FullyConnected, ScaleShift and Eltwise layers weights order from NCHW to NHWC.
- * Information for transposition is found from convolution/pooling input or output dimensions.
- * Convolution weights are transposed in finalizeConvolution1DPrimitive() method (gna_graph_compiler.cpp).
- * They are transposed for the both, NCHW and NHWC models since MO always stores them in NCHW layout.
- */
-DECL_PASS(TransposeWeightsFromNCHWToNHWC);
-
 /**
 * @brief fuse FullyConnected and Eltwise layers, also in case there is a Reshape between them having input with only
 * one dimension > 1
--- a/src/plugins/intel_gna/src/transformations/decompose_mvn.cpp
+++ b/src/plugins/intel_gna/src/transformations/decompose_mvn.cpp
@@ -26,7 +26,7 @@ struct MVNData {
    size_t W;
    size_t num_parts;
    float eps;
-    op::MVNEpsMode eps_mode;
+    ov::op::MVNEpsMode eps_mode;
    bool normalize_variance;
    element::Type element_type;
    std::string name;
@@ -138,7 +138,7 @@ static std::shared_ptr<Node> NormalizeVariance(const std::shared_ptr<opset8::MVN
                                                                       CoordinateDiff{0, 0},
                                                                       CoordinateDiff{0, 0},
                                                                       Strides{1, 1},
-                                                                       op::PadType::VALID);
+                                                                       ov::op::PadType::VALID);
    transposed_avg_conv_3->set_friendly_name(mvn_data.name + "_Avg3");
    auto avg_conv_3 =
        std::make_shared<opset8::Transpose>(transposed_avg_conv_3,
@@ -156,7 +156,7 @@ static std::shared_ptr<Node> NormalizeVariance(const std::shared_ptr<opset8::MVN
                                                                       CoordinateDiff{0, 0},
                                                                       CoordinateDiff{0, 0},
                                                                       Strides{1, 1},
-                                                                       op::PadType::VALID);
+                                                                       ov::op::PadType::VALID);
    transposed_avg_conv_4->set_friendly_name(mvn_data.name + "_Avg4");
    auto avg_conv_4 =
        std::make_shared<opset8::Transpose>(transposed_avg_conv_4,
@@ -243,7 +243,7 @@ static void Decompose(const std::shared_ptr<opset8::MVN> mvn, const MVNData& mvn
                                                                       CoordinateDiff{0, 0},
                                                                       CoordinateDiff{0, 0},
                                                                       Strides{1, 1},
-                                                                       op::PadType::VALID);
+                                                                       ov::op::PadType::VALID);
    transposed_avg_conv_1->set_friendly_name(mvn_data.name + "_Avg1");
    auto avg_conv_1 =
        std::make_shared<opset8::Transpose>(transposed_avg_conv_1,
@@ -261,7 +261,7 @@ static void Decompose(const std::shared_ptr<opset8::MVN> mvn, const MVNData& mvn
                                                                       CoordinateDiff{0, 0},
                                                                       CoordinateDiff{0, 0},
                                                                       Strides{1, 1},
-                                                                       op::PadType::VALID);
+                                                                       ov::op::PadType::VALID);
    transposed_avg_conv_2->set_friendly_name(mvn_data.name + "_Avg2");
    auto avg_conv_2 =
        std::make_shared<opset8::Transpose>(transposed_avg_conv_2,
--- a/src/plugins/intel_gna/src/transformations/fuse_conv_bias_activation.cpp
+++ b/src/plugins/intel_gna/src/transformations/fuse_conv_bias_activation.cpp
@@ -1,4 +1,4 @@
-// Copyright (C) 2018-2023 Intel Corporation
+// Copyright (C) 2023 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 //
 #include "fuse_conv_bias_activation.hpp"
--- a/src/plugins/intel_gna/src/transformations/gather_sinking.cpp
+++ b/src/plugins/intel_gna/src/transformations/gather_sinking.cpp
@@ -12,6 +12,9 @@
 #include "openvino/pass/pattern/op/wrap_type.hpp"
 #include "transformations/gather_sinking_binary.hpp"
 #include "transformations/gather_sinking_fuse.hpp"
+#include "transformations/gather_sinking_matmul.hpp"
+#include "transformations/gather_sinking_reshape.hpp"
+#include "transformations/gather_sinking_split.hpp"
 #include "transformations/gather_sinking_transpose_reshape.hpp"
 #include "transformations/gather_sinking_unary.hpp"

@@ -25,6 +28,7 @@ GatherSinkingGeneralForward::GatherSinkingGeneralForward() {
    add_matcher<GatherSinkingUnaryForward>();
    add_matcher<GatherSinkingBinaryForward>();
    add_matcher<GatherSinkingTransposeReshapeForward>();
+    add_matcher<GatherSinkingMatmulForward>();
    add_matcher<GatherSinkingFuse>();
 }

@@ -33,6 +37,9 @@ GatherSinkingGeneralBackward::GatherSinkingGeneralBackward() {
    add_matcher<GatherSinkingUnaryBackward>();
    add_matcher<GatherSinkingBinaryBackward>();
    add_matcher<GatherSinkingTransposeReshapeBackward>();
+    add_matcher<GatherSinkingReshapeBackward>();
+    add_matcher<GatherSinkingSplitBackward>();
+    add_matcher<GatherSinkingMatmulBackward>();
    add_matcher<GatherSinkingFuse>();
 }

--- a/src/plugins/intel_gna/src/transformations/handle_transposes_around_matmul.cpp
+++ b/src/plugins/intel_gna/src/transformations/handle_transposes_around_matmul.cpp
@@ -14,9 +14,11 @@
 #include <openvino/cc/ngraph/itt.hpp>

 #include "backend/gna_limitations.hpp"
+#include "common/graph_utils.hpp"

 using namespace ov::intel_gna::pass;
 using namespace ov::intel_gna::limitations;
+using namespace ov::intel_gna::graph_utils;

 namespace {

@@ -171,7 +173,7 @@ HandleTransposeBeforeMatMul::HandleTransposeBeforeMatMul() {
        auto iter = pattern_map.find(fq);
        if (iter != pattern_map.end() || (iter = pattern_map.find(constant)) != pattern_map.end()) {
            auto prev_node = iter->second.get_node_shared_ptr();
-            if (Limitations::is_transpose_2d(prev_node->get_output_shape(0))) {
+            if (is_shape_2d(prev_node->get_output_shape(0))) {
                InsertTranspose(prev_node, prev_node->get_friendly_name(), true);
            }
        }
--- a/src/plugins/intel_gna/src/transformations/markup_fusable_transpose.cpp
+++ b/src/plugins/intel_gna/src/transformations/markup_fusable_transpose.cpp
@@ -34,7 +34,7 @@ bool MarkupFusableTranspose::run_on_model(const std::shared_ptr<ngraph::Function

    for (auto& node : f->get_ordered_ops()) {
        if (!std::dynamic_pointer_cast<ngraph::opset9::Convolution>(node) &&
-            !std::dynamic_pointer_cast<ngraph::op::ConvolutionIE>(node)) {
+            !std::dynamic_pointer_cast<ov::intel_gna::op::GNAConvolution>(node)) {
            continue;
        }
        auto in_dims = node->input(0).get_shape();
--- a/src/plugins/intel_gna/src/transformations/replace_gna_nhwc_layers.cpp
+++ b/src/plugins/intel_gna/src/transformations/replace_gna_nhwc_layers.cpp
@@ -47,7 +47,7 @@ ov::Shape make_transpose_order_nchw2nhwc(size_t shape_size) {
    ov::Shape shape(shape_size);
    std::iota(shape.begin(), shape.end(), 0);

-    for (int i = 1; i < shape.size() - 1; ++i)
+    for (size_t i = 1; i < shape.size() - 1; ++i)
        shape[i] = shape[i + 1];

    *(shape.end() - 1) = 1;
--- a/src/plugins/intel_gna/src/transformations/reshape_transpose_substitute.cpp
+++ b/src/plugins/intel_gna/src/transformations/reshape_transpose_substitute.cpp
@@ -0,0 +1,162 @@
+// Copyright (C) 2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/reshape_transpose_substitute.hpp"
+
+#include <openvino/cc/ngraph/itt.hpp>
+#include <transformations/utils/utils.hpp>
+#include <utility>
+
+#include "openvino/opsets/opset12.hpp"
+#include "openvino/pass/pattern/op/wrap_type.hpp"
+
+using namespace ov;
+using namespace ov::opset12;
+using namespace ov::pass::pattern;
+using namespace ov::op::util;
+using namespace ov::intel_gna::pass;
+
+namespace {
+
+/**
+ * @brief Checks if output_shape is just permutation of input_shape dimensions
+ */
+bool is_shape_permutation(const Shape& input_shape, const Shape& output_shape) {
+    std::unordered_map<Shape::value_type, size_t> values_count;
+    for (const auto& dim : input_shape) {
+        auto it = values_count.find(dim);
+        if (it != values_count.end()) {
+            ++it->second;
+            continue;
+        }
+        values_count[dim] = 1;
+    }
+    for (const auto& dim : output_shape) {
+        auto it = values_count.find(dim);
+        if (it != values_count.end()) {
+            --it->second;
+            if (!it->second) {
+                values_count.erase(it);
+            }
+            continue;
+        }
+        return false;
+    }
+    return values_count.empty();
+}
+
+/**
+ * @brief Checks if all shape dimensions have different values
+ */
+bool are_dims_unique(const Shape& shape) {
+    std::unordered_set<Shape::value_type> dims;
+    for (const auto& dim : shape) {
+        if (dims.find(dim) != dims.end())
+            return false;
+        dims.insert(dim);
+    }
+    return true;
+}
+
+AxisVector get_unique_shapes_transpose_order(const Shape& input_shape, const Shape& output_shape) {
+    std::unordered_map<Shape::value_type, size_t> input_shape_items;
+    for (size_t i = 0; i < input_shape.size(); ++i) {
+        input_shape_items[input_shape[i]] = i;
+    }
+
+    AxisVector order;
+    order.reserve(output_shape.size());
+    for (const auto& output_dim : output_shape) {
+        order.push_back(input_shape_items.find(output_dim)->second);
+    }
+
+    return order;
+}
+
+/**
+ * @brief Changes shape according to specified order
+ */
+Shape apply_permutation(const Shape& shape, const AxisVector& order) {
+    Shape transposed_shape;
+    transposed_shape.reserve(shape.size());
+    for (const auto& position : order) {
+        transposed_shape.push_back(shape[position]);
+    }
+    return transposed_shape;
+}
+
+AxisVector find_suitable_transpose_order(const Shape& input_shape,
+                                         const Shape& output_shape,
+                                         const std::vector<AxisVector>& orders) {
+    for (const auto& order : orders) {
+        const Shape transposed_shape = apply_permutation(input_shape, order);
+        if ((transposed_shape.size() == output_shape.size()) &&
+            std::equal(transposed_shape.begin(), transposed_shape.end(), output_shape.begin()))
+            return order;
+    }
+
+    return {};
+}
+
+AxisVector find_suitable_transpose_order(const Shape& input_shape, const Shape& output_shape) {
+    static std::vector<AxisVector> orders_4d = {AxisVector{0, 2, 3, 1}, AxisVector{0, 3, 1, 2}};
+    static std::vector<AxisVector> orders_3d = {AxisVector{1, 2, 0}, AxisVector{2, 0, 1}};
+
+    switch (input_shape.size()) {
+    case 4:
+        return find_suitable_transpose_order(input_shape, output_shape, orders_4d);
+    case 3:
+        return find_suitable_transpose_order(input_shape, output_shape, orders_3d);
+    case 2:
+        return AxisVector{1, 0};
+    default:
+        return {};
+    }
+    return {};
+}
+
+/**
+ * @brief Finds applicable transpose order to get output_shape from input_shape if
+ * it is possible
+ * Returns empty AxisVector on unsupported case
+ */
+AxisVector get_transpose_order(const Shape& input_shape, const Shape& output_shape) {
+    if (are_dims_unique(input_shape))
+        return get_unique_shapes_transpose_order(input_shape, output_shape);
+    return find_suitable_transpose_order(input_shape, output_shape);
+}
+
+}  // namespace
+
+ReshapeTransposeSubstitute::ReshapeTransposeSubstitute() {
+    MATCHER_SCOPE(ReshapeTransposeSubstitute);
+
+    auto reshape_label = wrap_type<Reshape>({any_input(), any_input()}, [](const Output<Node>& output) {
+        if (!has_static_shape()(output))
+            return false;
+        const Node* node = output.get_node();
+        return is_shape_permutation(node->get_input_shape(0), node->get_output_shape(0));
+    });
+
+    ov::matcher_pass_callback matcher_pass_callback = [=](Matcher& m) {
+        const auto& pattern_to_output = m.get_pattern_value_map();
+        auto reshape = as_type_ptr<Reshape>(pattern_to_output.at(reshape_label).get_node_shared_ptr());
+
+        const AxisVector transpose_order =
+            get_transpose_order(reshape->get_input_shape(0), reshape->get_output_shape(0));
+        if (transpose_order.empty())
+            return false;
+
+        auto new_transpose_const =
+            std::make_shared<Constant>(element::i64, Shape{transpose_order.size()}, transpose_order);
+        auto new_transpose = std::make_shared<Transpose>(reshape->input_value(0), new_transpose_const);
+
+        replace_node_update_name(reshape, new_transpose);
+
+        return true;
+    };
+
+    auto m = std::make_shared<Matcher>(reshape_label, matcher_name);
+    register_matcher(m, matcher_pass_callback);
+}
--- a/src/plugins/intel_gna/src/transformations/reshape_transpose_substitute.hpp
+++ b/src/plugins/intel_gna/src/transformations/reshape_transpose_substitute.hpp
@@ -0,0 +1,34 @@
+// Copyright (C) 2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/pass/graph_rewrite.hpp"
+#include "transformations_visibility.hpp"
+
+namespace ov {
+namespace intel_gna {
+namespace pass {
+
+/**
+ * @brief Substitute Reshape with Transpose if it possible.
+ * It is usefull with Tranpose/Gather sinking transformations to simplify Transpose movement.
+ * TransposeSinking is not possible to use with Reshapes (except some cases when Reshape is Squeeze/Unsqueeze).
+ * GatherSinking is much harder to implement for all existed layer types.
+ *
+ * any layer          any layer
+ *    |                  |
+ *  Reshape   =>      Transpose
+ *    |                  |
+ * any layer          any layer
+ */
+class ReshapeTransposeSubstitute : public ov::pass::MatcherPass {
+public:
+    OPENVINO_RTTI("ReshapeTransposeSubstitute", "0");
+    ReshapeTransposeSubstitute();
+};
+
+}  // namespace pass
+}  // namespace intel_gna
+}  // namespace ov
--- a/src/plugins/intel_gna/src/transformations/rotate_inputs.cpp
+++ b/src/plugins/intel_gna/src/transformations/rotate_inputs.cpp
@@ -0,0 +1,79 @@
+// Copyright (C) 2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/rotate_inputs.hpp"
+
+#include "common/graph_utils.hpp"
+#include "openvino/cc/ngraph/itt.hpp"
+#include "openvino/opsets/opset11.hpp"
+#include "openvino/pass/pattern/op/wrap_type.hpp"
+#include "ops/gna_convolution.hpp"
+
+using namespace ov::opset11;
+using namespace ov::pass;
+using namespace ov::intel_gna::pass;
+using namespace ov::intel_gna::graph_utils;
+
+namespace {
+inline bool is_skip_operation(const std::shared_ptr<ov::Node>& node) {
+    return (std::dynamic_pointer_cast<Reshape>(node) != nullptr ||
+            std::dynamic_pointer_cast<Transpose>(node) != nullptr ||
+            std::dynamic_pointer_cast<Squeeze>(node) != nullptr ||
+            std::dynamic_pointer_cast<Unsqueeze>(node) != nullptr ||
+            std::dynamic_pointer_cast<FakeQuantize>(node) != nullptr) &&
+           has_n_consumers(node, 1);
+}
+}  // namespace
+
+InsertConvolutionTransposeHW::InsertConvolutionTransposeHW() {
+    MATCHER_SCOPE(InsertConvolutionTransposeHW);
+
+    const auto conv_pattern = pattern::wrap_type<op::GNAConvolution>(
+        {pattern::any_input(), pattern::any_input()},
+        [](const ov::Output<ov::Node>& node) {
+            std::shared_ptr<op::GNAConvolution> conv =
+                std::dynamic_pointer_cast<op::GNAConvolution>(node.get_node_shared_ptr());
+            helper::ConvData conv_data;
+            helper::GetConvData(conv, conv_data);
+            return gna_convolution_layer::should_transpose_h_w(static_cast<uint32_t>(conv_data.input_height),
+                                                               static_cast<uint32_t>(conv_data.filter_height),
+                                                               static_cast<uint32_t>(conv_data.input_channel_count),
+                                                               static_cast<uint32_t>(conv_data.filter_stride_height));
+        });
+
+    ov::matcher_pass_callback callback = [=](pattern::Matcher& m) {
+        const auto& pattern_map = m.get_pattern_value_map();
+
+        // auto param_node = pattern_map.at(param_pattern).get_node_shared_ptr();
+        auto conv_node = pattern_map.at(conv_pattern).get_node_shared_ptr();
+
+        std::shared_ptr<ov::Node> target_node =
+            graph_utils::get_prev_node_skipping_certain(conv_node->get_input_node_shared_ptr(0), is_skip_operation);
+        std::shared_ptr<Parameter> param_node = std::dynamic_pointer_cast<Parameter>(target_node);
+
+        if (!param_node) {
+            return false;
+        }
+
+        // transpose all convolution inputs
+        for (const auto& conv_input : conv_node->inputs()) {
+            // Transpose H and W (NHWC -> NWHC)
+            ov::AxisVector tr_axis = {0, 2, 1, 3};
+            auto transpose_const = std::make_shared<Constant>(ov::element::i8, ov::Shape{tr_axis.size()}, tr_axis);
+            auto transpose = std::make_shared<Transpose>(conv_input.get_source_output(), transpose_const);
+
+            // Reshape out
+            ov::Shape shape_out = conv_input.get_shape();
+            auto reshape_out_const =
+                std::make_shared<Constant>(ov::element::i32, ov::Shape{shape_out.size()}, shape_out);
+            auto reshape_out = std::make_shared<Reshape>(transpose, reshape_out_const, false);
+
+            conv_input.replace_source_output(reshape_out);
+        }
+        return true;
+    };
+
+    auto m = std::make_shared<pattern::Matcher>(conv_pattern, matcher_name);
+    this->register_matcher(m, callback);
+}
--- a/src/plugins/intel_gna/src/transformations/rotate_inputs.hpp
+++ b/src/plugins/intel_gna/src/transformations/rotate_inputs.hpp
@@ -0,0 +1,41 @@
+// Copyright (C) 2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/pass/graph_rewrite.hpp"
+
+namespace ov {
+namespace intel_gna {
+namespace pass {
+
+/**
+ * @brief Transpose convolution inputs (HW->WH)
+ * when it is directly connected to network inputs
+ *
+ * Searches for next pattern
+ *     Any input layer
+ *           |
+ *   Reshape/FQ/Squeeze/Usqueeze
+ *           |
+ *        Convolution
+ *
+ *    And transforms to
+ *     Any input layer
+ *           |
+ *       Transpose
+ *           |
+ *    Reshape/FQ/Squeeze/Usqueeze
+ *           |
+ *        Convolution
+ */
+class InsertConvolutionTransposeHW : public ov::pass::MatcherPass {
+public:
+    OPENVINO_RTTI("InsertConvolutionTransposeHW", "0");
+    InsertConvolutionTransposeHW();
+};
+
+}  // namespace pass
+}  // namespace intel_gna
+}  // namespace ov
--- a/src/plugins/intel_gna/src/transformations/ts_concat_forward.cpp
+++ b/src/plugins/intel_gna/src/transformations/ts_concat_forward.cpp
@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 //

-#include "transformations/ts_concat.hpp"
+#include "transformations/ts_concat_forward.hpp"

 #include <openvino/cc/ngraph/itt.hpp>

@@ -23,7 +23,6 @@ using namespace ov::intel_gna::pass::helper;
 using namespace ov::intel_gna::limitations;

 namespace {
-
 bool is_concat_sinked(const Output<Node>& output) {
    auto concat_node = ov::as_type_ptr<Concat>(output.get_node_shared_ptr());

@@ -41,7 +40,6 @@ bool is_concat_sinked(const Output<Node>& output) {

    return false;
 }
-
 }  // namespace

 TSConcatForward::TSConcatForward() {
--- a/src/plugins/intel_gna/src/transformations/ts_concat_forward.hpp
+++ b/src/plugins/intel_gna/src/transformations/ts_concat_forward.hpp
--- a/src/plugins/intel_gna/src/transformations/ts_split_backward.cpp
+++ b/src/plugins/intel_gna/src/transformations/ts_split_backward.cpp
@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 //

-#include "transformations/ts_split.hpp"
+#include "transformations/ts_split_backward.hpp"

 #include <openvino/cc/ngraph/itt.hpp>

--- a/src/plugins/intel_gna/src/transformations/ts_split_backward.hpp
+++ b/src/plugins/intel_gna/src/transformations/ts_split_backward.hpp
--- a/src/plugins/intel_gna/src/transformations/utils/transformation_helper.cpp
+++ b/src/plugins/intel_gna/src/transformations/utils/transformation_helper.cpp
@@ -6,17 +6,20 @@

 #include "log/debug.hpp"
 #include "openvino/core/rt_info.hpp"
-#include "openvino/opsets/opset10.hpp"
+#include "openvino/opsets/opset12.hpp"
 #include "openvino/pass/pattern/op/wrap_type.hpp"
+#include "ops/gna_convolution.hpp"
+#include "ops/gna_max_pool.hpp"
+#include "transformations/rt_info/transpose_sinking_attr.hpp"
+
+using namespace ov::opset12;

 namespace ov {
 namespace intel_gna {
 namespace pass {
 namespace helper {

-using namespace ov::opset10;
-
-void GetConvData(std::shared_ptr<ngraph::op::ConvolutionIE> conv, ConvData& conv_data) {
+void GetConvData(std::shared_ptr<ngraph::opset7::Convolution> conv, ConvData& conv_data) {
    OPENVINO_ASSERT(conv);
    conv_data.output_height = conv->get_output_shape(0)[2];
    conv_data.output_width = conv->get_output_shape(0)[3];
@@ -40,17 +43,17 @@ void GetConvData(std::shared_ptr<ngraph::op::ConvolutionIE> conv, ConvData& conv
    conv_data.element_type = conv->get_element_type();
 }

-void GetConvData(std::shared_ptr<Convolution> conv, ConvData& conv_data) {
+void GetConvData(std::shared_ptr<ov::intel_gna::op::GNAConvolution> conv, ConvData& conv_data) {
    OPENVINO_ASSERT(conv);
    conv_data.output_height = conv->get_output_shape(0)[2];
    conv_data.output_width = conv->get_output_shape(0)[3];
-    conv_data.input_channel_count = conv->input_value(0).get_shape()[1];
-    conv_data.input_height = conv->input_value(0).get_shape()[2];
-    conv_data.input_width = conv->input_value(0).get_shape()[3];
+    conv_data.input_channel_count = conv->input_value(0).get_shape()[3];
+    conv_data.input_height = conv->input_value(0).get_shape()[1];
+    conv_data.input_width = conv->input_value(0).get_shape()[2];
    conv_data.filter_count = conv->input_value(1).get_shape()[0];
-    conv_data.filter_channel_count = conv->input_value(1).get_shape()[1];
-    conv_data.filter_height = conv->input_value(1).get_shape()[2];
-    conv_data.filter_width = conv->input_value(1).get_shape()[3];
+    conv_data.filter_channel_count = conv->input_value(1).get_shape()[3];
+    conv_data.filter_height = conv->input_value(1).get_shape()[1];
+    conv_data.filter_width = conv->input_value(1).get_shape()[2];
    conv_data.filter_dilation_height = conv->get_dilations()[0];
    conv_data.filter_dilation_width = conv->get_dilations()[1];
    conv_data.filter_stride_height = conv->get_strides()[0];
@@ -180,6 +183,75 @@ void swap_names(std::shared_ptr<ov::Node> node1, std::shared_ptr<ov::Node> node2
    swap_output_names(node1->output(0), node2->output(0));
 }

+ov::AxisVector reverse_transpose_order(const ov::AxisVector& axis_order) {
+    ov::AxisVector out(axis_order.size());
+    for (size_t i = 0; i < axis_order.size(); i++) {
+        out.at(axis_order[i]) = i;
+    }
+    return out;
+}
+
+ov::NodeVector find_input_transposes(const std::shared_ptr<const ov::Node>& node) {
+    ov::NodeVector transposes;
+    for (size_t input_idx = 0; input_idx < node->get_input_size(); ++input_idx) {
+        auto input_node = node->get_input_node_shared_ptr(input_idx);
+        auto transpose_node = ov::as_type_ptr<Transpose>(input_node);
+        if (transpose_node)
+            transposes.push_back(transpose_node);
+    }
+
+    return transposes;
+}
+
+void mark_input_transposes_as_nosinking(std::shared_ptr<const ov::Node> node) {
+    for (const auto& input : find_input_transposes(node))
+        ov::mark_as_no_sinking_node(input);
+}
+
+TransposeInfo get_first_input_transpose(const std::shared_ptr<const ov::Node>& node) {
+    for (size_t input_idx = 0; input_idx < node->get_input_size(); ++input_idx) {
+        std::shared_ptr<Node> input_node = node->get_input_node_shared_ptr(input_idx);
+        auto transpose_node = as_type_ptr<Transpose>(input_node);
+        if (!transpose_node)
+            continue;
+        auto constant_node = as_type_ptr<Constant>(transpose_node->input_value(1).get_node_shared_ptr());
+        if (!constant_node)
+            continue;
+        {
+            TransposeInfo input_info;
+            input_info.transpose = transpose_node;
+            input_info.transpose_const = constant_node;
+            return input_info;
+        }
+    }
+
+    return {};
+}
+
+TransposeInfo get_first_output_transpose(const std::shared_ptr<const ov::Node>& node) {
+    for (size_t i = 0; i < node->get_output_size(); ++i) {
+        for (const auto& input : node->output(i).get_target_inputs()) {
+            Node* node = input.get_node();
+            if (!dynamic_cast<Transpose*>(node))
+                continue;
+            auto transpose_node = as_type_ptr<Transpose>(node->shared_from_this());
+            if (!transpose_node)
+                continue;
+            auto constant_node = as_type_ptr<Constant>(transpose_node->input_value(1).get_node_shared_ptr());
+            if (!constant_node)
+                continue;
+            {
+                TransposeInfo input_info;
+                input_info.transpose = transpose_node;
+                input_info.transpose_const = constant_node;
+                return input_info;
+            }
+        }
+    }
+
+    return {};
+}
+
 }  // namespace helper
 }  // namespace pass
 }  // namespace intel_gna
--- a/src/plugins/intel_gna/src/transformations/utils/transformation_helper.hpp
+++ b/src/plugins/intel_gna/src/transformations/utils/transformation_helper.hpp
@@ -7,6 +7,9 @@
 #include <legacy/ngraph_ops/convolution_ie.hpp>
 #include <ngraph/opsets/opset7.hpp>

+#include "openvino/opsets/opset12.hpp"
+#include "ops/gna_convolution.hpp"
+
 namespace ov {
 namespace intel_gna {
 namespace pass {
@@ -44,12 +47,12 @@ struct ConvData {
 void GetConvData(std::shared_ptr<ngraph::opset7::Convolution> conv, ConvData& conv_data);

 /**
- * @brief gets all legacy convolution related data into a struct for further processing
- * @param conv legacy convolution node to get data of
+ * @brief gets all convolution related data into a struct for further processing
+ * @param conv GNA custom convolution node to get data of
 * @param conv_data convolution data structure to put data into
 * @return void
 */
-void GetConvData(std::shared_ptr<ngraph::op::ConvolutionIE> conv, ConvData& conv_data);
+void GetConvData(std::shared_ptr<ov::intel_gna::op::GNAConvolution> conv, ConvData& conv_data);

 /**
 * @brief ngraph matcher predicate fusing existing predicates for consumers count and rank of a layer
@@ -116,6 +119,41 @@ void swap_friendly_names(std::shared_ptr<ov::Node>, std::shared_ptr<ov::Node>);
 */
 void swap_names(std::shared_ptr<ov::Node>, std::shared_ptr<ov::Node>);

+/**
+ * @brief Reverses axis order. Final result will be such an order, that together
+ * with initial order will be {0, 1, 2, ...}
+ */
+ov::AxisVector reverse_transpose_order(const ov::AxisVector& axis_order);
+
+/**
+ * @brief Finds all input node transposes
+ */
+ov::NodeVector find_input_transposes(const std::shared_ptr<const ov::Node>& node);
+
+/**
+ * @brief Marks all input transposes with flag NoSinking
+ */
+void mark_input_transposes_as_nosinking(std::shared_ptr<const ov::Node> node);
+
+struct TransposeInfo {
+    std::shared_ptr<ov::opset12::Transpose> transpose;
+    std::shared_ptr<ov::opset12::Constant> transpose_const;
+
+    bool isEmpty() const {
+        return !transpose || !transpose_const;
+    }
+};
+
+/**
+ * @brief Finds first input node transpose
+ */
+TransposeInfo get_first_input_transpose(const std::shared_ptr<const ov::Node>& node);
+
+/**
+ * @brief Finds first output node transpose
+ */
+TransposeInfo get_first_output_transpose(const std::shared_ptr<const ov::Node>& node);
+
 }  // namespace helper
 }  // namespace pass
 }  // namespace intel_gna
--- a/src/plugins/intel_gna/tests/deprecated/unit/engines/gna/i16_quantisation_test.cpp
+++ b/src/plugins/intel_gna/tests/deprecated/unit/engines/gna/i16_quantisation_test.cpp
@@ -273,32 +273,6 @@ TEST_F(I16QuantisationTest, EltwiseSumm_onlyOneIdentityInsertion) {
        .once();
 }

-TEST_F(I16QuantisationTest, canDetectLeakyRelu) {
-    assert_that()
-        .onInferModel(TFLeakyReluModel())
-        .inNotCompactMode()
-        .withGNAConfig(GNA_CONFIG_KEY(SCALE_FACTOR), 1.0f)
-        .gna()
-        .propagate_forward()
-        .called_with()
-        .pwl_inserted_into_nnet();
-}
-
-TEST_F(I16QuantisationTest, MaxPool_followedAfterActivation) {
-    assert_that()
-        .onInferModel(maxpoolAfterRelu())
-        .inNotCompactMode()
-        .withGNAConfig(GNA_CONFIG_KEY(SCALE_FACTOR), 1.0f)
-        .gna()
-        .propagate_forward()
-        .called_with()
-        .convolution_inserted_into_nnet()
-        .And()
-        .pwl_inserted_into_nnet()
-        .And()
-        .max_pooling_inserted_into_nnet();
-}
-
 TEST_F(I16QuantisationTest, EltwiseMull_willInsertTwoIdentities) {
    assert_that()
        .onInferModel(eltwiseMulModel())
--- a/src/plugins/intel_gna/tests/functional/pass_tests/concat_transpose.cpp
+++ b/src/plugins/intel_gna/tests/functional/pass_tests/concat_transpose.cpp
@@ -0,0 +1,137 @@
+// Copyright (C) 2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+#include <map>
+#include <numeric>
+#include <ostream>
+#include <string>
+#include <vector>
+
+#include "common_test_utils/common_utils.hpp"
+#include "ngraph_functions/builders.hpp"
+#include "openvino/opsets/opset12.hpp"
+#include "shared_test_classes/base/layer_test_utils.hpp"
+#include "shared_test_classes/base/ov_subgraph.hpp"
+
+using namespace ov::opset12;
+using namespace ov::test;
+
+namespace TransposesConcatTests {
+
+typedef std::tuple<ov::element::Type,                   // Net precision
+                   std::string,                         // Device name
+                   std::map<std::string, std::string>,  // Configuration
+                   std::map<std::string, std::string>   // Additional Configuration
+                   >
+    TransposesConcatTestParamsSet;
+
+class TransposesConcatTest : public testing::WithParamInterface<TransposesConcatTestParamsSet>,
+                             virtual public SubgraphBaseTest {
+public:
+    static std::string get_test_case_name(const testing::TestParamInfo<TransposesConcatTestParamsSet>& obj) {
+        ov::element::Type net_type, in_type, out_type;
+        std::string target_device;
+        std::map<std::string, std::string> conf, conf_ext;
+
+        std::tie(net_type, target_device, conf, conf_ext) = obj.param;
+        for (auto& conf_item : conf_ext) {
+            conf[conf_item.first] = conf_item.second;
+        }
+
+        std::ostringstream result;
+        result << "netPRC=" << net_type << "_";
+        result << "trgDev=" << target_device;
+        for (auto const& conf_i : conf) {
+            result << "_configItem=" << conf_i.first.c_str() << "_" << conf_i.second.c_str();
+        }
+        return result.str();
+    }
+
+protected:
+    void SetUp() override {
+        abs_threshold = std::numeric_limits<int32_t>::max();
+        rel_threshold = std::numeric_limits<int32_t>::max();
+        std::map<std::string, std::string> conf, conf_ext;
+        std::tie(m_net_type, targetDevice, conf, conf_ext) = this->GetParam();
+
+        std::vector<InputShape> input_shapes = static_shapes_to_test_representation({{10, 1}});
+        configuration.insert(conf.begin(), conf.end());
+        for (auto& conf_item : conf_ext) {
+            configuration[conf_item.first] = conf_item.second;
+        }
+        init_input_shapes(input_shapes);
+    }
+
+    void init_test_model() {
+        std::vector<std::vector<size_t>> input_shapes = {{10, 1}};
+
+        auto params = ngraph::builder::makeParams(m_net_type, input_shapes);
+
+        std::vector<size_t> shape_1 = {10, 128};
+        std::vector<size_t> shape_2 = {10, 192};
+        std::vector<size_t> shape_3 = {10, 256};
+
+        std::shared_ptr<ov::Node> input_node = params[0];
+
+        std::vector<float> mulConstWeights(1 * 576);
+        std::iota(mulConstWeights.begin(), mulConstWeights.end(), 0.1f);
+        auto constMul1 = ngraph::builder::makeConstant<float>(m_net_type, ov::Shape{1, 576}, mulConstWeights);
+        auto matmul1 = std::make_shared<ov::opset10::MatMul>(input_node, constMul1, false, false);
+
+        auto split_axis = std::make_shared<Constant>(ov::element::u8, ov::Shape{1}, std::vector<uint8_t>{1});
+        auto split_slices =
+            std::make_shared<Constant>(ov::element::u32, ov::Shape{3}, std::vector<uint32_t>{128, 192, 256});
+        auto split_node = std::make_shared<VariadicSplit>(matmul1, split_axis, split_slices);
+
+        std::vector<size_t> transpose_order = {1, 0};
+        auto transpose_const_1 =
+            std::make_shared<Constant>(ov::element::u8, ov::Shape{transpose_order.size()}, transpose_order);
+        auto transpose_const_2 =
+            std::make_shared<Constant>(ov::element::u8, ov::Shape{transpose_order.size()}, transpose_order);
+        auto transpose_const_3 =
+            std::make_shared<Constant>(ov::element::u8, ov::Shape{transpose_order.size()}, transpose_order);
+
+        auto transpose_node_1 = std::make_shared<Transpose>(split_node->output(0), transpose_const_1);
+        auto transpose_node_2 = std::make_shared<Transpose>(split_node->output(1), transpose_const_2);
+        auto transpose_node_3 = std::make_shared<Transpose>(split_node->output(2), transpose_const_3);
+
+        const int axis = 0;
+        auto concat_node =
+            std::make_shared<Concat>(ov::OutputVector{transpose_node_1, transpose_node_2, transpose_node_3}, axis);
+
+        std::vector<size_t> reshape_pattern = {1, 2, 5, 576};
+        auto reshape_const =
+            std::make_shared<Constant>(ov::element::u16, ov::Shape{reshape_pattern.size()}, reshape_pattern);
+        auto reshape_node = std::make_shared<Reshape>(concat_node, reshape_const, false);
+
+        ov::ResultVector results{std::make_shared<Result>(reshape_node)};
+        function = std::make_shared<ov::Model>(results, params, "concat");
+    }
+
+    ov::element::Type m_net_type;
+    std::vector<size_t> m_input_shape;
+};
+
+TEST_P(TransposesConcatTest, CompareWithRefs) {
+    init_test_model();
+    run();
+}
+
+std::vector<std::map<std::string, std::string>> configs = {{{"GNA_DEVICE_MODE", "GNA_SW_EXACT"}}};
+
+std::vector<std::map<std::string, std::string>> target_configs = {{{"GNA_DEVICE_MODE", "GNA_SW_FP32"}},
+                                                                  {{"GNA_EXEC_TARGET", "GNA_TARGET_2_0"}},
+                                                                  {{"GNA_EXEC_TARGET", "GNA_TARGET_3_0"}},
+                                                                  {{"GNA_EXEC_TARGET", "GNA_TARGET_3_5"}}};
+
+const ov::element::TypeVector input_precisions = {ov::element::f32};
+
+INSTANTIATE_TEST_SUITE_P(smoke_transposes_concat,
+                         TransposesConcatTest,
+                         ::testing::Combine(::testing::ValuesIn(input_precisions),
+                                            ::testing::Values(CommonTestUtils::DEVICE_GNA),
+                                            ::testing::ValuesIn(configs),
+                                            ::testing::ValuesIn(target_configs)),
+                         TransposesConcatTest::get_test_case_name);
+
+}  // namespace TransposesConcatTests
--- a/src/plugins/intel_gna/tests/functional/shared_tests_instances/single_layer_tests/convolution.cpp
+++ b/src/plugins/intel_gna/tests/functional/shared_tests_instances/single_layer_tests/convolution.cpp
@@ -108,7 +108,7 @@ void ConvolutionLayerTestFixture::SetUp() {

    auto filter_size = std::accumulate(std::begin(kernel), std::end(kernel), 1, std::multiplies<size_t>());
    filter_weights =
-        CommonTestUtils::generate_float_numbers(convOutChannels * inputShape[1] * filter_size, -0.5f, 0.5f);
+        CommonTestUtils::generate_float_numbers(convOutChannels * inputShape[1] * filter_size, -0.1f, 0.1f);

    auto conv = std::dynamic_pointer_cast<Convolution>(ngraph::builder::makeConvolution(paramOuts[0],
                                                                                        ngPrc,
--- a/src/plugins/intel_gna/tests/unit/gna_memory_compact_test.cpp
+++ b/src/plugins/intel_gna/tests/unit/gna_memory_compact_test.cpp
@@ -310,11 +310,11 @@ class GNAMemoryOrderTest : public ::testing::Test {};
 TEST_F(GNAMemoryOrderTest, orderingFusedLayersActivation) {
    auto plugin = GNAPluginTested();

-    ov::Shape input_shape = {1, 8, 20, 16};
+    ov::Shape input_shape = {1, 16, 20, 16};
    ov::Strides strides = {1, 1};
    ov::Strides dilations = {1, 1};
    ov::CoordinateDiff pad_begin(0, 0), pad_end(0, 0);
-    auto weights = ngraph::builder::makeConstant<float>(ov::element::f32, {8, 8, 1, 1}, {1.f});
+    auto weights = ngraph::builder::makeConstant<float>(ov::element::f32, {8, 16, 1, 1}, {1.f});

    auto input = std::make_shared<ngraph::opset8::Parameter>(ov::element::f32, input_shape);
    auto conv = std::make_shared<ngraph::opset8::Convolution>(input, weights, strides, pad_begin, pad_end, dilations);
@@ -332,11 +332,11 @@ TEST_F(GNAMemoryOrderTest, orderingFusedLayersActivation) {
 TEST_F(GNAMemoryOrderTest, orderingFusedLayersMaxPool) {
    auto plugin = GNAPluginTested();

-    ov::Shape input_shape = {1, 8, 20, 16};
+    ov::Shape input_shape = {1, 16, 20, 16};
    ov::Strides strides = {1, 1};
    ov::Strides dilations = {1, 1};
    ov::CoordinateDiff pad_begin(0, 0), pad_end(0, 0);
-    auto weights = ngraph::builder::makeConstant<float>(ov::element::f32, {8, 8, 1, 1}, {1.f});
+    auto weights = ngraph::builder::makeConstant<float>(ov::element::f32, {8, 16, 1, 1}, {1.f});

    auto input = std::make_shared<ngraph::opset8::Parameter>(ov::element::f32, input_shape);
    auto conv = std::make_shared<ngraph::opset8::Convolution>(input, weights, strides, pad_begin, pad_end, dilations);
@@ -361,11 +361,11 @@ TEST_F(GNAMemoryOrderTest, orderingFusedLayersMaxPool) {
 TEST_F(GNAMemoryOrderTest, orderingFusedLayersActivationMaxPool) {
    auto plugin = GNAPluginTested();

-    ov::Shape input_shape = {1, 8, 20, 16};
+    ov::Shape input_shape = {1, 16, 20, 16};
    ov::Strides strides = {1, 1};
    ov::Strides dilations = {1, 1};
    ov::CoordinateDiff pad_begin(0, 0), pad_end(0, 0);
-    auto weights = ngraph::builder::makeConstant<float>(ov::element::f32, {8, 8, 1, 1}, {1.f});
+    auto weights = ngraph::builder::makeConstant<float>(ov::element::f32, {8, 16, 1, 1}, {1.f});

    auto input = std::make_shared<ngraph::opset8::Parameter>(ov::element::f32, input_shape);
    auto conv = std::make_shared<ngraph::opset8::Convolution>(input, weights, strides, pad_begin, pad_end, dilations);
--- a/src/plugins/intel_gna/tests/unit/transformations/gather_sinking_test.cpp
+++ b/src/plugins/intel_gna/tests/unit/transformations/gather_sinking_test.cpp
@@ -0,0 +1,130 @@
+// Copyright (C) 2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "transformations/gather_sinking.hpp"
+
+#include <ngraph/function.hpp>
+#include <ngraph/pass/manager.hpp>
+#include <openvino/opsets/opset10.hpp>
+#include <ops/gna_convolution.hpp>
+#include <ops/gna_max_pool.hpp>
+#include <transformations/init_node_info.hpp>
+
+#include "common_test_utils/ngraph_test_utils.hpp"
+#include "gtest/gtest.h"
+#include "transformations/gather_sinking_matmul.hpp"
+#include "transformations/gather_sinking_reshape.hpp"
+#include "transformations/gather_sinking_split.hpp"
+#include "transformations/replace_gna_nhwc_layers.hpp"
+
+using namespace ov;
+using namespace ov::opset10;
+
+TEST(TransposeNCHW, Convolution) {
+    std::shared_ptr<Model> function;
+    {
+        auto input_params = std::make_shared<Parameter>(element::Type_t::f32, Shape{1, 1, 41, 1});
+        auto kernel = Constant::create(ov::element::f32, {4, 1, 3, 1}, {1});
+
+        auto convolution = std::make_shared<Convolution>(input_params,
+                                                         kernel,
+                                                         Strides{2, 1},
+                                                         CoordinateDiff{0, 0},
+                                                         CoordinateDiff{0, 0},
+                                                         Strides{1, 1});
+
+        const auto result = std::make_shared<Result>(convolution);
+        function = std::make_shared<Model>(OutputVector{result}, ParameterVector{input_params});
+    }
+
+    std::shared_ptr<Model> orig_function = function->clone();
+    ov::pass::Manager manager;
+    manager.register_pass<ov::pass::InitNodeInfo>();
+    manager.register_pass<ov::intel_gna::pass::SubstituteGNAConvolution>();
+    manager.run_passes(function);
+    ASSERT_NO_THROW(check_rt_info(function));
+
+    std::shared_ptr<Model> reference_function;
+    {
+        auto input_params = std::make_shared<Parameter>(element::Type_t::f32, Shape{1, 1, 41, 1});
+
+        auto transpose_before_const = Constant::create(element::i32, Shape{4}, {0, 2, 3, 1});
+
+        auto transpose_before = std::make_shared<Transpose>(input_params, transpose_before_const);
+
+        auto kernel = Constant::create(ov::element::f32, {4, 1, 3, 1}, {1});
+
+        auto transpose_conv_const = Constant::create(element::i32, Shape{4}, {0, 2, 3, 1});
+
+        auto transpose_conv_before = std::make_shared<Transpose>(input_params, transpose_conv_const);
+
+        auto transpose_conv_constant = std::make_shared<Transpose>(kernel, transpose_conv_const);
+
+        auto convolution = std::make_shared<ov::intel_gna::op::GNAConvolution>(transpose_before,
+                                                                               transpose_conv_constant,
+                                                                               Strides{2, 1},
+                                                                               CoordinateDiff{0, 0},
+                                                                               CoordinateDiff{0, 0},
+                                                                               Strides{1, 1});
+
+        auto transpose_after_const = Constant::create(element::i32, Shape{4}, {0, 3, 1, 2});
+
+        auto transpose_after = std::make_shared<Transpose>(convolution, transpose_after_const);
+
+        const auto result = std::make_shared<Result>(transpose_after);
+        reference_function = std::make_shared<Model>(OutputVector{result}, ParameterVector{input_params});
+    }
+
+    const FunctionsComparator func_comparator =
+        FunctionsComparator::with_default().enable(FunctionsComparator::ATTRIBUTES);
+    const FunctionsComparator::Result result = func_comparator(function, reference_function);
+    ASSERT_TRUE(result.valid) << result.message;
+}
+
+TEST(TransposeNCHW, MaxPool) {
+    std::shared_ptr<Model> function;
+    {
+        auto input_params = std::make_shared<Parameter>(element::Type_t::f32, Shape{1, 1, 41, 1});
+
+        auto max_pool =
+            std::make_shared<ov::op::v1::MaxPool>(input_params, Strides{2, 1}, Shape{0, 0}, Shape{0, 0}, Shape{4, 1});
+
+        const auto result = std::make_shared<Result>(max_pool);
+        function = std::make_shared<Model>(OutputVector{result}, ParameterVector{input_params});
+    }
+
+    std::shared_ptr<Model> orig_function = function->clone();
+    ov::pass::Manager manager;
+    manager.register_pass<ov::pass::InitNodeInfo>();
+    manager.register_pass<ov::intel_gna::pass::SubstituteGNAMaxPool>();
+    manager.run_passes(function);
+    ASSERT_NO_THROW(check_rt_info(function));
+
+    std::shared_ptr<Model> reference_function;
+    {
+        auto input_params = std::make_shared<Parameter>(element::Type_t::f32, Shape{1, 1, 41, 1});
+
+        auto transpose_before_const = Constant::create(element::i32, Shape{4}, {0, 2, 3, 1});
+
+        auto transpose_before = std::make_shared<Transpose>(input_params, transpose_before_const);
+
+        auto max_pool = std::make_shared<ov::intel_gna::op::GNAMaxPool>(transpose_before,
+                                                                        Strides{2, 1},
+                                                                        Shape{0, 0},
+                                                                        Shape{0, 0},
+                                                                        Shape{4, 1});
+
+        auto transpose_after_const = Constant::create(element::i32, Shape{4}, {0, 3, 1, 2});
+
+        auto transpose_after = std::make_shared<Transpose>(max_pool, transpose_after_const);
+
+        const auto result = std::make_shared<Result>(transpose_after);
+        reference_function = std::make_shared<Model>(OutputVector{result}, ParameterVector{input_params});
+    }
+
+    const FunctionsComparator func_comparator =
+        FunctionsComparator::with_default().enable(FunctionsComparator::ATTRIBUTES);
+    const FunctionsComparator::Result result = func_comparator(function, reference_function);
+    ASSERT_TRUE(result.valid) << result.message;
+}
--- a/src/plugins/intel_gna/tests/unit/transformations/gather_sinking_ts_split_test.cpp
+++ b/src/plugins/intel_gna/tests/unit/transformations/gather_sinking_ts_split_test.cpp
@@ -8,7 +8,7 @@
 #include "openvino/opsets/opset12.hpp"
 #include "openvino/pass/manager.hpp"
 #include "transformations/init_node_info.hpp"
-#include "transformations/ts_split.hpp"
+#include "transformations/ts_split_backward.hpp"

 using namespace ov;
 using namespace ov::opset12;
--- a/src/tests/functional/shared_test_classes/src/single_layer/convolution.cpp
+++ b/src/tests/functional/shared_test_classes/src/single_layer/convolution.cpp
@@ -58,7 +58,7 @@ void ConvolutionLayerTest::SetUp() {
    if (targetDevice == CommonTestUtils::DEVICE_GNA) {
        auto filter_size = std::accumulate(std::begin(kernel), std::end(kernel), 1, std::multiplies<size_t>());
        filter_weights = CommonTestUtils::generate_float_numbers(convOutChannels * inputShape[1] * filter_size,
-                                                                 -0.5f, 0.5f);
+                                                                 -0.1f, 0.1f);
    }
    auto conv = std::dynamic_pointer_cast<ngraph::opset1::Convolution>(
            ngraph::builder::makeConvolution(paramOuts[0], ngPrc, kernel, stride, padBegin,
--- a/src/tests/functional/shared_test_classes/src/subgraph/input_conv.cpp
+++ b/src/tests/functional/shared_test_classes/src/subgraph/input_conv.cpp
@@ -56,7 +56,7 @@ void InputConvTest::SetUp() {
        std::vector<float> res;
        for (std::size_t i = 0; i < out_channels; ++i) {
            for (std::size_t j = 0; j < kernel_size; ++j) {
-                j == 0 ? res.emplace_back(1.0f) : res.emplace_back(0.0f);
+                j == 0 ? res.emplace_back(0.2f) : res.emplace_back(0.0f);
            }
        }