IntenseWebs/openvino
3
0 Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
13,725 Commits 38 Branches 54 Tags
032ac898e21c4008fb8c8dc2da5778c1b3ab4afe
Commit Graph
100 Commits
This Branch
This Branch
All Branches
Author SHA1 Message Date
Sergey Shlyapnikov 032ac898e2 [GPU] Apply is_non_decompression_multiply() callback only for compressed models (#21719) 2023-12-18 12:20:20 +00:00
Sergey Shlyapnikov a04d59d779 [GPU] Gemm tiled opt add dynamic padding support (#21640) 2023-12-14 09:38:16 +04:00
Sergey Shlyapnikov 15421001fc [GPU] Add missed override keyword in typed_primitive_onednn_impl struct (#21593) 2023-12-12 14:01:34 +00:00
Sergey Shlyapnikov 8d77a52e81 [GPU] Add callback function for MarkDequantizationSubgraph pass to skip non FC-related subgraphs (#21552) 2023-12-11 08:59:46 +04:00
Sergey Shlyapnikov 0ad0a6bfaa [GPU] Improve fake alignment logic and fix missed zero point value (#21530) 
* [GPU] Improve fake alignment logic and fix missed zero point value

* Add tests
 2023-12-08 12:20:35 +00:00
Sergey ShlyapnikovandKim, Eddy 152b4dfc0d [GPU] Add SLM support for FC bf tiled kernel (#21435) 
* [GPU] Add SLM support for FC bf tiled kernel

* Fix unaligned IFM leftovers processing in case of compressed weights and add decompression scale post op support

* added FullyConnected_bf_tiled::GetUpdateDispatchDataFunc

* updated FullyConnected_bf_tiled::GetUpdateDispatchDataFunc for two types of kernels

---------

Co-authored-by: Kim, Eddy <eddy.kim@intel.com>
 2023-12-06 18:18:01 +01:00
Sergey Shlyapnikov 009951d969 [GPU] Allow to set empty tensor for inner network of condition primitive (#21415) 2023-12-01 21:21:21 +00:00
Sergey Shlyapnikov e23704ef5e [GPU] Update FC batch blocking for INT4 weights (#21228) 2023-11-23 13:24:14 +04:00
Sergey Shlyapnikov af34891750 [GPU] Add more flexible check if user's buffer is USM Host and can be reused (for iGPUs only) (#21018) 2023-11-16 09:46:10 +04:00
Sergey Shlyapnikov c360de0f07 [GPU] Optimize Condition operation by better integrating its subnetwork primitives with the main network (#21062) 
* [GPU] Optimize Condition operation by better integrating its subnetwork primitives with the main network

* Reuse main network's stream in sub-networks
 2023-11-15 08:47:36 +04:00
Sergey Shlyapnikov 306137f86b [GPU] Make ShapePredictor instance unique for each InferRequest instead of the cldnn::network (#21019) 2023-11-13 08:52:07 +04:00
Sergey Shlyapnikov 26cab79988 [GPU] Improve OOOQ synchronization between ShapeOf subgraph's CPU impls and GPU kernels (#20595) 
* [GPU] Add ITT counters for primitive inst

* [GPU] Improve OOOQ synchronization between ShapeOf subgraph's CPU impls and GPU kernels
 2023-10-24 18:49:41 +04:00
Sergey Shlyapnikov 6c396fac9d [GPU] Deferred events deallocation and tuning for FC bf_tiled kernel (#20519) 2023-10-18 16:28:27 +04:00
Sergey ShlyapnikovandVladimir Paramuzov 6e97b91a77 [GPU] UsmHostTensor implementation (#20518) 
Co-authored-by: Vladimir Paramuzov <vladimir.paramuzov@intel.com>
 2023-10-18 16:12:15 +04:00
Sergey Shlyapnikov f655d2177d [GPU] Fix scatter_nd_update output paddings handling (#20447) 2023-10-16 13:52:45 +04:00
Sergey Shlyapnikov 146ca36f8f [GPU] Minor debug improvements: add suffix to shape agnostic kernels names and show batch_hash for impl_types::any (#20275) 2023-10-16 13:47:57 +04:00
Sergey Shlyapnikov ec644b9a73 [GPU] Fix device tensors reallocation in case of host user's tensors (#20306) 2023-10-11 10:47:33 +04:00
Sergey Shlyapnikov b9ff53a5a4 [GPU] Fix runtime reorders optimization and add runtime memory dependencies (#20202) 2023-10-04 12:57:02 +04:00
Sergey Shlyapnikov e2501a67d2 [GPU] Fix os_is_yx_osv16_isv4 weights format matching (#20132) 2023-10-02 09:02:10 +04:00
Sergey Shlyapnikov f9678a285c [GPU] Do not use usm_host memory buffers for PVC as a device inputs (#19767) 2023-09-28 13:53:51 +04:00
Sergey Shlyapnikov 886be26c0b [GPU] Add handling of unsupported simd8 for PVC (#20093) 2023-09-28 13:39:41 +04:00
Sergey Shlyapnikov 4eb9c57424 [GPU] Add input feature leftovers processing for fully_connected_gpu_bs_f_bsv16_af8_vload kernel (#19650) 2023-09-07 13:20:11 +04:00
Sergey Shlyapnikov 7df8d1ca2d [GPU] Add per iteration performance profiling mode (#18637) 2023-08-22 12:37:36 +02:00
Sergey Shlyapnikov d13ae31a61 [GPU] Allow to use infinity value as a Pad's fill value (#19201) 2023-08-15 14:31:49 +00:00
Sergey Shlyapnikov 86c4c6785d [GPU] Don't invalidate users after reorder insertion (#19137) 2023-08-14 14:16:54 +04:00
Sergey Shlyapnikov 32a6a31de2 [GPU] Prevent Conv's input data type changing at reorder_inputs pass (#19042) 2023-08-10 10:59:32 +04:00
Sergey Shlyapnikov d91d72c89c [GPU] RandomUniform new shape inference for dynamism support (#19087) 2023-08-10 09:43:08 +04:00
Sergey Shlyapnikov e7565eda3e [GPU] Added paddings processing for dynamic convolutions primitives; enabled convolution osv16 optimized shape agnostic kernel (#19001) 2023-08-08 14:57:54 +02:00
Sergey Shlyapnikov 2bd49cc1f8 [GPU] Add extra flush() command for dynamic models and OOO queue (#18819) 2023-07-27 16:15:37 +04:00
Sergey Shlyapnikov 7f917446c2 [GPU] Force ov::enable_profiling if OV_GPU_DumpProfilingData debug option is used (#18687) 2023-07-26 13:32:30 +04:00
Sergey Shlyapnikov 42237bcbdb [GPU] Optimize update_shape synchronization for OOOQ and prohibit memory reuse for shape_of-subgraph nodes (#18742) 2023-07-24 11:34:19 -07:00
Sergey Shlyapnikov 3958f771da [GPU] Add oneDNN primitives profiling support (#17785) 
* [GPU] Add oneDNN primitives profiling support

* [GPU] Add stream.wait() method to prevent caches flushing and other possible impacts of finish() call

* Add comment for wait() usage
 2023-07-17 19:17:31 +04:00
Sergey Shlyapnikov b7935bb869 [GPU] Fix data races in RegisterFactory() and LogHelper::LogHelper() functions (#18521) 2023-07-14 09:06:10 +04:00
Sergey Shlyapnikov feab0ab75e [GPU] Memory preallocation optimization (#18181) 
* [GPU] Use real layout for cpu impls instead of memory's

* [GPU] Add memory tracking and pre allocation mechanism

* Tests and minor code refactoring

* Apply review comments

* Remove unused include
 2023-07-13 07:26:17 +02:00
Sergey Shlyapnikov e631f65a9b [GPU] Fix in-order queue synchronization issue related to OCL/OneDNN impls interaction with CPU impls (#17976) 2023-06-14 10:15:04 +09:00
Sergey Shlyapnikov 70e0caca4f [GPU] Fix dynamic padding processing of static dimension (#17978) 2023-06-12 08:39:42 +04:00
Sergey Shlyapnikov 58d79aa3a6 [GPU] Add shape_of subgraphs markup and initial cpu implementations (#17762) 
* [GPU] Add shape of subgraphs markup and initial cpu implementations for some of primitives

* Apply review comments

* Exclude eltwise with boolean mode types from shape of subgraphs and fix leftovers
 2023-06-08 13:46:21 +04:00
Sergey Shlyapnikov db8d23231a [GPU] Change priority of CPU implementations (#17829) 2023-06-05 11:21:26 +04:00
Sergey Shlyapnikov 5afbd4cf92 [GPU] Remove clFinish call from USM memory lock function (#17830) 2023-06-02 16:17:05 +04:00
Sergey ShlyapnikovandVladimir Paramuzov 808647dfb3 [GPU] Weights reorders primitive cachcing (#16638) 
Co-authored-by: Vladimir Paramuzov <vladimir.paramuzov@intel.com>
 2023-05-15 09:57:04 +04:00
Sergey Shlyapnikov 3b8bb1fb17 [GPU] Use BFS processing order for out_of_order queue (#17250) 2023-05-02 15:14:38 +04:00
Sergey Shlyapnikov 0a5975bdfa [GPU] Add real kernels' execution timings collection for DumpProfilingData debug option (#15797) 2023-04-25 14:33:08 +04:00
Sergey Shlyapnikov 997f60f1c3 [GPU] Fix shape_of shape inference optimization (#16863) 2023-04-12 15:44:34 +04:00
Sergey Shlyapnikov 2075dcb7c3 [GPU] Fix Interpolate assert (#16806) 2023-04-10 12:29:01 +04:00
Sergey Shlyapnikov 48dee7c30a [GPU] Fix missed weights params update (#16815) 2023-04-10 10:28:06 +04:00
Sergey Shlyapnikov bb93bfd90f [GPU] Add clDNN shape agnostic kernels usage as an initial impls for dGPU (#16018) 
* [GPU] Add clDNN shape agnostic kernels usage as an initial impls for dGPU

* [GPU] Use layout as a key of weights cache, implement logic for weights cache capacity calculation based on available memory
 2023-03-31 13:05:59 +04:00
Sergey Shlyapnikov 6c766a81b5 [GPU] Treat warnings C4267 as errors for Windows (#16345) 2023-03-28 22:56:47 +00:00
Sergey Shlyapnikov ce67ac09d3 [GPU] Disable OneDNN primitive cache (#16525) 2023-03-26 23:29:47 +04:00
Sergey Shlyapnikov e434c320f5 [GPU] Update tuning params of shape agnostic version of fully_connected_bf_tiled kernel for dGPUs (#16482) 2023-03-24 09:08:09 +04:00
Sergey Shlyapnikov 7b649c4150 [GPU] Fix reset_execution() method with wait option enabled for in_order queue type (#15562) 2023-02-08 09:18:46 +04:00
Sergey Shlyapnikov e003bf3af7 [GPU] Shape agnostic FC opt tiled kernel (#15396) 2023-02-06 12:17:55 +04:00
Sergey Shlyapnikov cd48d76009 [GPU] Limit legacy fusions usage in Convolution kernels (#15465) 2023-02-06 12:08:22 +04:00
Sergey Shlyapnikov 6c22f06d07 [GPU] Perform memory transfer from usm_host to usm_device for dGPU only (#15263) 
* [GPU] Perform memory transfer from usm_host to usm_device only for dGPU

* [GPU] Allocate new memory buffer for biases fusions to avoid original buffer modification since it may be used by other primitives
 2023-01-26 10:44:38 +04:00
Sergey Shlyapnikov 19fde3fee6 [GPU] Peak memory usage optimizations (#14799) 2023-01-02 09:51:19 +04:00
Sergey Shlyapnikov 901e7982d6 [GPU] Fix Experimental Detectron ROI Feature Extractor pyramid levels indexing and Experimental Detectron DO in case of class_agnostic_box_regression option (#14700) 2022-12-21 15:11:33 +04:00
Sergey Shlyapnikov f02c663a1d [GPU] Fix eltwise fusions in case of LT_ALIGNED_READ and vec_size>1 (#14275) 2022-12-20 12:13:15 +04:00
Sergey Shlyapnikov 9f481e8eaa [GPU] Fix Pad Primitive shape inference (#14282) 2022-12-13 16:59:55 +04:00
Sergey Shlyapnikov 637f1cdf8f [GPU] Fix lockable memory allocation (#14134) 
* [GPU] Fix lockable memory allocation

* Apply review comments

* [GPU] Replace usm handle sharing with independent blob creation

* Prohibit all other remote blob types

* [GPU] Fix for nullptr impl
 2022-11-25 12:07:47 +04:00
Sergey Shlyapnikov 06cfa2abc4 [GPU] Support of surface input reordering to buffer (#13223) 
* [GPU] Add surface input reordering

* [GPU] Add batched nv12->bfyx conversion reorder fusion
 2022-11-24 16:20:38 +04:00
Sergey Shlyapnikov d464d079b2 [GPU] Add blocking option for cldnn::memory::copy_*() functions (#14012) 2022-11-24 10:37:11 +04:00
Sergey Shlyapnikov be277ab977 [GPU] Move device memory allocation from preprocessing stage to runtime for usm memory type (#13697) 
* [GPU] Move device memory allocation from preprocessing stage to runtime to fix incorrect behaviour in case of tensor swapping

* [GPU] Apply review comments

* Replace source of dynamic layout cheking
 2022-11-17 21:49:28 +03:00
Sergey Shlyapnikov a1b496e051 [GPU] Default context for each device (#13658) 
* [GPU] Add default context for each device

* [GPU] Replace operator== overloading with independent function
 2022-11-02 11:04:56 +04:00
Sergey Shlyapnikov 6c6b3666a4 [GPU] Add additional condition for VA surface buffers detection (#13691) 2022-10-28 18:19:02 +03:00
Sergey Shlyapnikov 5fea4c3fc3 [GPU] Convert activation's slope buffer data type for OneDNN fusions (#13444) 2022-10-14 13:58:35 +04:00
Sergey Shlyapnikov 0b8f1f8c00 [GPU] Optimize reorder between permute and first convolution (#13459) 2022-10-14 13:57:58 +04:00
Sergey Shlyapnikov bdffe5a054 [GPU] Fix reshape shape calculation for MatMul operation (#13312) 2022-10-04 11:15:07 +04:00
Sergey Shlyapnikov 33cda223a5 [GPU] Fix performance regression caused by surfaces_lock::create() function move (#13124) 2022-09-22 10:06:09 +04:00
Sergey Shlyapnikov 2fc59bf3d2 [GPU] GPU Plugin OVInferRequestIOTensorTest and SqueezeUnsqueezeLayerTest tests fixes (#12830) 2022-09-01 09:04:27 +04:00
Sergey Shlyapnikov 30b872150a [GPU] Fix Interpolate axes values alignment with scales (#12634) 2022-08-25 09:39:30 +04:00
Sergey Shlyapnikov bf2ae06f29 [GPU] Shared memory optimization for network::execute_impl() call (#12631) 2022-08-24 19:11:21 +04:00
Sergey Shlyapnikov 7d2c88b296 Fix path for find_package of ITT counters in case of defined INTEL_VTUNE_DIR variable (#12584) 2022-08-16 16:45:29 +04:00
Sergey Shlyapnikov d8237cd97c [GPU] Align Reduce primitive with nGraph's parameters (#12341) 2022-08-16 13:53:26 +04:00
Sergey Shlyapnikov a711ae87a5 [GPU] Align Interpolate primitive with nGraph's parameters (#12342) 
* [GPU] Align Interpolate primitive with nGraph's parameters
 2022-08-16 13:52:46 +04:00
Sergey Shlyapnikov 79e3272237 [GPU] Update eltwise calc_output_layout function and prevent output layouts invalidation after adding reorders for weights (#11073) 2022-03-31 13:15:47 +03:00
Sergey Shlyapnikov f2af1ef88a [GPU] Update memory location to __local in GPU Detection Output (#11209) 
* [GPU] Update memory location to __local in GPU Detection Output

* Replace hardcoded stack size with JIT constant
 2022-03-31 13:13:58 +03:00
Sergey Shlyapnikov 24a74672f6 [GPU] Fix remote blobs tests (#11349) 2022-03-31 09:04:46 +03:00
Sergey Shlyapnikov 5dc3da926c [GPU] Add fusion through feature (#9674) 
* [GPU] Add fuse through feature

* Apply review comments


Tasks marked as not passed:
* https://dev.azure.com/openvinoci/dldt/_build/results?buildId=313769&view=results
* https://dev.azure.com/openvinoci/dldt/_build/results?buildId=313766&view=results
 2022-03-30 16:43:32 +03:00
Sergey Shlyapnikov cd703580b6 [GPU] Host time optimizations for in order queue (#11255) 
* [GPU] Host time optimizations

* Fix failed fusings_gpu/permute_eltwise_loop.basic/* tests
 2022-03-30 10:53:53 +03:00
Sergey Shlyapnikov 782ef6b42e [GPU] Add estimation of required memory for CAFFE_OPT_2 stage of GPU Detection Output (#11001) 2022-03-21 12:06:07 +03:00
Sergey Shlyapnikov 6c6aa8fa95 [GPU] Fix RemoteBlob lock() and ulock() behaviour in case of multiple threads (#10685) 
* [GPU] Fix RemoteBlob lock() and ulock() behaviour in case of multiple threads and add tests
 2022-03-01 15:15:04 +03:00
Sergey Shlyapnikov e906b3581f [GPU] Replace handle_permute optimization pass with proper Reorder adding instead of Permute primitive (#10569) 2022-02-24 16:41:43 +03:00
Sergey ShlyapnikovandMaxim Shevtsov 03566b4e4d [GPU] Fix outputs blobs allocation for U16/I16 data types (#10180) 
* [GPU] Fix outputs blobs allocation for U16/I16 data types

* [GPU] Add U32, U64, FP64 data types support; add information about legacy fused activations to .info file

* Update auto_batch.cpp

fixed u64 inputs for the auto-batching

Co-authored-by: Maxim Shevtsov <maxim.y.shevtsov@intel.com>
 2022-02-11 14:08:36 +03:00
Sergey Shlyapnikov bc21e52912 [GPU] Fix FC 3D input size propagation and bias fusion (#10249) 2022-02-10 16:16:12 +03:00
Sergey Shlyapnikov 8f0e974ee6 [GPU] Add new properties and fix bechmark_app (#10149) 2022-02-09 09:18:54 +03:00
Sergey Shlyapnikov ccf4f4e420 [GPU] Update config api 2.0 (#9649) 2022-02-03 13:04:36 +03:00
Sergey Shlyapnikov 0af8916d1d [GPU] Fix dependencies list in fused primitive descriptor (#9941) 2022-01-27 11:05:20 +03:00
Sergey Shlyapnikov 89868df352 [GPU] Fix eltwise's inputs indexes calculation and order of dimensions (#9778) 2022-01-20 11:15:40 +03:00
Sergey Shlyapnikov 0a225c2d84 [GPU] Fix fusions indexes calculation in b_fs_yx_fsv32 pooling kernel (#9756) 2022-01-19 11:18:53 +03:00
Sergey Shlyapnikov bd12648ec0 [GPU] Fix users invalidation and output layout recalculation in graph dump mode (#9439) 2022-01-13 12:29:43 +03:00
Sergey Shlyapnikov 95d86eb2bf [GPU] Add parallel quantizes optimization (#9370) 2021-12-27 09:47:20 +03:00
Sergey Shlyapnikov 507a498269 [GPU] Add OneDNN post ops description in graph dump mode (#9371) 2021-12-23 09:58:35 +03:00
Sergey Shlyapnikov 71f3e2da10 [GPU] Fix fused ops reshaping for OneDNN FC (#9217) 2021-12-17 13:26:59 +03:00
Sergey Shlyapnikov 2514c0ef38 [GPU] Add gemm_tiled_opt i8/u8 output support (#9202) 2021-12-16 15:20:28 +03:00
Sergey ShlyapnikovandKim,SungEun 1177d2b282 [GPU] Change FQ output for first Convolution (#9200) 
* update onednn_gpu

* [GPU] Add bs_fs_yx_bsv8_fsv4 format

Co-authored-by: Kim,SungEun <sungeun.kim@intel.com>
 2021-12-15 13:15:13 +03:00
Sergey Shlyapnikov 5817f191c3 [GPU] Fix quantization type detection of GEMM and FC primitives (#9113) 2021-12-10 09:45:15 +03:00
Sergey Shlyapnikov 7002dd4317 [GPU] Add b_fs_yx_fsv32 format support for border primitive and update tuning logic of convolution_b_fs_zyx_fsv16_imad kernel for Gen12HP (#9112) 2021-12-09 18:43:46 +03:00
Sergey Shlyapnikov fa1d9867f2 [GPU] Fix incorrect output result in primitives with two outputs (#8902) 2021-12-01 19:23:38 +03:00
Sergey Shlyapnikov 497aaf9b46 [GPU] Relax batch size limitaions for DO GPU implementation usage for some of Gen12HP dGPUs (#8801) 2021-11-29 11:41:39 +09:00
Sergey Shlyapnikov e76fc14ae1 [GPU] Use b_fs_yx_fsv16 format for OneDNN convolutins in case of FP32 output (#8808) 2021-11-27 15:08:06 +03:00
Sergey Shlyapnikov c10a981f1b [GPU] Do not use OneDNN for FP32 pipeline (#8737) 2021-11-25 10:58:37 +03:00