Code

oneDNN 2.5 migration#121

Code

Open

dmitry-gorokhov wants to merge 63 commits intov2.5_for_ie_masteropenvinotoolkit/oneDNN:v2.5_for_ie_masterfrom v2.5_for_ie_master_local_changesopenvinotoolkit/oneDNN:v2.5_for_ie_master_local_changesCopy head branch name to clipboard

Commits

Commits on Feb 20, 2022

• jit: avx512: conv: Fix missed ur_w iteration

  Show description for 8109bd8
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  8109bd8

  Copy full SHA for 8109bd8

• Enable jit sse41 NxN convolution for grayscale input

  

  dmitrygo

  committed
  e248b50

  Copy full SHA for e248b50

• Support of strided blobs for [de]convolution and simple reorder

  Show description for 1c6c17d
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  1c6c17d

  Copy full SHA for 1c6c17d

• Updated sse41 jit convolutions to support padded channels

  

  dmitrygo

  committed
  94dbed3

  Copy full SHA for 94dbed3

• Introduced Depthwise and Quantization post ops

  Show description for a707196
  

  dmitrygo

  committed
  a707196

  Copy full SHA for a707196

• Pooling pads like Caffe

  Show description for d31c5c4
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  d31c5c4

  Copy full SHA for d31c5c4

• TBB_AUTO was enabled

  

  alexey-varyzgin

  authored and

  dmitrygo

  committed
  21d958b

  Copy full SHA for 21d958b

• Add API function dnnl_memory_set_data_handle_no_pads_proc

  Show description for 23dacbf
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  23dacbf

  Copy full SHA for 23dacbf

• nchw_pooling dense fix

  

  alexey-varyzgin

  authored and

  dmitrygo

  committed
  d22cb71

  Copy full SHA for d22cb71

• Enabled BWD (JIT/GEMM) FP32/BF16 Convoltions + Depthwise post ops fusings

  

  dmitrygo

  committed
  b17c0cc

  Copy full SHA for b17c0cc

• Fixes for MKLDNN to enable LTO

  

  ilya-lavrenov

  authored and

  dmitrygo

  committed
  d83837b

  Copy full SHA for d83837b

• [MSVC] Enabling SIMD functionality for VS2019

  

  mandrono

  authored and

  dmitrygo

  committed
  8e0bbf7

  Copy full SHA for 8e0bbf7

• Avoid usage of undefined macro

  Show description for 6a7071e
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  6a7071e

  Copy full SHA for 6a7071e

• Add several uni instruction wrappers into jit_generator

  Show description for d13149c
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  d13149c

  Copy full SHA for d13149c

• Fix ODR violation

  Show description for 11c81cd
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  11c81cd

  Copy full SHA for 11c81cd

• fix name matching with system strauct 'user' in llvm-android toolchain

  Show description for b625aec
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  b625aec

  Copy full SHA for b625aec

• Added JIT FP32/BF16 Softmax for arbitrary inner_size

  Show description for e333684
  

  dmitrygo

  committed
  e333684

  Copy full SHA for e333684

• Added support of hsigmoid, round_half_to_even, round_half_away_from_zero elementwise algorithms

  

  a-sidorova

  authored and

  dmitrygo

  committed
  31cd484

  Copy full SHA for 31cd484

• Limit applicability of is_1stconv logic for JIT FP32/BF16 AVX512 Convolution

  

  dmitrygo

  committed
  c74c508

  Copy full SHA for c74c508

• [WA] Removed kernel_outside_src condition on JIT FP32/BF16 Convolutions

  

  dmitrygo

  committed
  3b61547

  Copy full SHA for 3b61547

• Added custom vesrion of JIT DW FP32/BF16 Convolution with 5D input support

  

  dmitrygo

  committed
  83137dc

  Copy full SHA for 83137dc

• Asymmetric quntization for activations

  

  dmitrygo

  committed
  11141a1

  Copy full SHA for 11141a1

• Added 3D DW case support for JIT INT8 Convolutions

  

  dmitrygo

  committed
  7d34a00

  Copy full SHA for 7d34a00

• [WA] Disabled weights md transpose in FC to prevent perf degradations

  

  dmitrygo

  committed
  fad9dbe

  Copy full SHA for fad9dbe

• Dynamic batch support via context

  

  dmitrygo

  committed
  8d279a1

  Copy full SHA for 8d279a1

• Added JIT AVX512/AVX2 FP32 Planar Convolution implementation

  

  dmitrygo

  committed
  50043b9

  Copy full SHA for 50043b9

• Binary networks support

  

  dmitrygo

  committed
  ce960ad

  Copy full SHA for ce960ad

• Accommodating oneTBB (with hybrid cores support) that

  Show description for 64232aa
  

  myshevts

  authored and

  dmitrygo

  committed
  64232aa

  Copy full SHA for 64232aa

• NCHW pooling perfomance fixed in accordance with v0.21

  

  maxnick

  authored and

  dmitrygo

  committed
  daeb468

  Copy full SHA for daeb468

• [WA] Fixed fallback on ref conv in case exceeding scratchpad limit

  

  dmitrygo

  committed
  4592ba4

  Copy full SHA for 4592ba4

• Returned old behavior for fp32 avx2 1x1 conv with dw conv fusing

  

  antonvor

  authored and

  dmitrygo

  committed
  0c99230

  Copy full SHA for 0c99230

• Updated SoftPlus

  

  a-sidorova

  authored and

  dmitrygo

  committed
  dcd0abf

  Copy full SHA for dcd0abf

• Fixed warning fo undefined ITT_ARCH_IA64 (#52)

  

  ilya-lavrenov

  authored and

  dmitrygo

  committed
  9ccc627

  Copy full SHA for 9ccc627

• Disable reorder JIT if both inputs and outputs are batch-strided.

  

  IvanNovoselov

  authored and

  dmitrygo

  committed
  ccd5353

  Copy full SHA for ccd5353

• Include TBB headers as system

  Show description for 94c0a20
  

  AlexPeskov

  authored and

  dmitrygo

  committed
  94c0a20

  Copy full SHA for 94c0a20

• Fixed redifinition of tls model

  

  dmitrygo

  committed
  f1fd7e4

  Copy full SHA for f1fd7e4

• nspc layout support for convolutions

  

  maxnick

  authored and

  dmitrygo

  committed
  a8f7910

  Copy full SHA for a8f7910

• set scale = 1.f in case of signed input on platforms without vnni

  

  antonvor

  authored and

  dmitrygo

  committed
  597c659

  Copy full SHA for 597c659

• Enable direct copy primitives for u8 reorder

  

  IvanNovoselov

  authored and

  dmitrygo

  committed
  90b00ac

  Copy full SHA for 90b00ac

• Memory descriptor dynamism related changes

  Show description for de66d63
  

  maxnick

  authored and

  dmitrygo

  committed
  de66d63

  Copy full SHA for de66d63

• Added prelu as binary post op

  

  antonvor

  authored and

  dmitrygo

  committed
  1501344

  Copy full SHA for 1501344

• Depthwise and Quantization post ops for Gemm Convolutions

  

  antonvor

  authored and

  dmitrygo

  committed
  a7c1712

  Copy full SHA for a7c1712

• Perf fixes for Ref and NCHW Poolings

  

  antonvor

  authored and

  dmitrygo

  committed
  12e2bca

  Copy full SHA for 12e2bca

• perf fixes for quantization post ops

  

  antonvor

  authored and

  dmitrygo

  committed
  ab4c85f

  Copy full SHA for ab4c85f

• todo: fix assert(idx < max_idx)

  

  antonvor

  authored and

  dmitrygo

  committed
  80a89c2

  Copy full SHA for 80a89c2

• simple reorder: temporarily disabled zero padding

  

  antonvor

  authored and

  dmitrygo

  committed
  f5f763d

  Copy full SHA for f5f763d

• Fix possible data race when accessing global reorder list

  

  Egor Duplensky

  authored and

  dmitrygo

  committed
  98064bb

  Copy full SHA for 98064bb

• Brgemm implementation has perf degradation in RNN node

  

  alexey-varyzgin

  authored and

  dmitrygo

  committed
  49d6a78

  Copy full SHA for 49d6a78

• reverted old behavior with pdim_consistent check due to perf problems

  

  antonvor

  authored and

  dmitrygo

  committed
  994f9c5

  Copy full SHA for 994f9c5

• Renamed matmul kernel type: brg -> brgemm

  

  dmitrygo

  committed
  a5aa5bb

  Copy full SHA for a5aa5bb

• Fixed bias addition order in brgemm kernel

  

  dmitrygo

  committed
  a4ebe69

  Copy full SHA for a4ebe69

• [1D] Enlarge support

  

  alexey-varyzgin

  authored and

  dmitrygo

  committed
  9f776d1

  Copy full SHA for 9f776d1

• Update uni_ jit methods to avoid mixing vex and nonVEX instructions

  

  Egor Duplensky

  authored and

  dmitrygo

  committed
  8ac9ff7

  Copy full SHA for 8ac9ff7

• Quantization post op structure modified to reduce its complexity

  

  maxnick

  authored and

  dmitrygo

  committed
  26a09d4

  Copy full SHA for 26a09d4

• Hash utility functions were extracted to a separate module for reuse

  

  maxnick

  authored and

  dmitrygo

  committed
  298546e

  Copy full SHA for 298546e

• Desc similar_to routine consider start stride

  

  maxnick

  authored and

  dmitrygo

  committed
  c4d9021

  Copy full SHA for c4d9021

• Desc similar_to routine use stride cmp mask

  

  maxnick

  authored and

  dmitrygo

  committed
  81698b9

  Copy full SHA for 81698b9

• added some legacy parallel methods to fix perf issues

  Show description for f8d99a9
  

  antonvor

  authored and

  dmitrygo

  committed
  f8d99a9

  Copy full SHA for f8d99a9

• Migrate legacy post ops and zero points on runtime data pointers

  

  dmitrygo

  committed
  5a77d6e

  Copy full SHA for 5a77d6e

Commits on Feb 21, 2022

• Revert "all: remove mkldnn compatibility layer"

  Show description for 15c2025
  

  dmitrygo

  committed
  15c2025

  Copy full SHA for 15c2025

Commits on Mar 5, 2022

• Revert "reverted old behavior with pdim_consistent check due to perf problems"

  Show description for 22e2744
  

  dmitrygo

  committed
  22e2744

  Copy full SHA for 22e2744

Commits on Mar 14, 2022

• Fixed ODR violataion

  

  dmitrygo

  committed
  5ac2a40

  Copy full SHA for 5ac2a40

Commits on Apr 5, 2022

• equality_uni_xxx_for_sse_and_avx

  

  chenhu-wang

  authored and

  dmitrygo

  committed
  d55288f

  Copy full SHA for d55288f