oneDNN 2.5 migration#121

Open

dmitry-gorokhov wants to merge 63 commits intov2.5_for_ie_masterfrom

v2.5_for_ie_master_local_changes

Commits on Feb 20, 2022

jit: avx512: conv: Fix missed ur_w iteration

AlexPeskov
authored and
dmitrygo
committed
Enable jit sse41 NxN convolution for grayscale input
dmitrygo
committed
Support of strided blobs for [de]convolution and simple reorder

AlexPeskov
authored and
dmitrygo
committed
Updated sse41 jit convolutions to support padded channels
dmitrygo
committed
Introduced Depthwise and Quantization post ops
dmitrygo
committed
Pooling pads like Caffe

AlexPeskov
authored and
dmitrygo
committed
TBB_AUTO was enabled

alexey-varyzgin
authored and
dmitrygo
committed
Add API function dnnl_memory_set_data_handle_no_pads_proc

AlexPeskov
authored and
dmitrygo
committed
nchw_pooling dense fix

alexey-varyzgin
authored and
dmitrygo
committed
Enabled BWD (JIT/GEMM) FP32/BF16 Convoltions + Depthwise post ops fusings
dmitrygo
committed
Fixes for MKLDNN to enable LTO

ilya-lavrenov
authored and
dmitrygo
committed
[MSVC] Enabling SIMD functionality for VS2019

mandrono
authored and
dmitrygo
committed
Avoid usage of undefined macro

AlexPeskov
authored and
dmitrygo
committed
Add several uni instruction wrappers into jit_generator

AlexPeskov
authored and
dmitrygo
committed
Fix ODR violation

AlexPeskov
authored and
dmitrygo
committed
fix name matching with system strauct 'user' in llvm-android toolchain

AlexPeskov
authored and
dmitrygo
committed
Added JIT FP32/BF16 Softmax for arbitrary inner_size
dmitrygo
committed
Added support of hsigmoid, round_half_to_even, round_half_away_from_zero elementwise algorithms

a-sidorova
authored and
dmitrygo
committed
Limit applicability of is_1stconv logic for JIT FP32/BF16 AVX512 Convolution
dmitrygo
committed
[WA] Removed kernel_outside_src condition on JIT FP32/BF16 Convolutions
dmitrygo
committed
Added custom vesrion of JIT DW FP32/BF16 Convolution with 5D input support
dmitrygo
committed
Asymmetric quntization for activations
dmitrygo
committed
Added 3D DW case support for JIT INT8 Convolutions
dmitrygo
committed
[WA] Disabled weights md transpose in FC to prevent perf degradations
dmitrygo
committed
Dynamic batch support via context
dmitrygo
committed
Added JIT AVX512/AVX2 FP32 Planar Convolution implementation
dmitrygo
committed
Binary networks support
dmitrygo
committed
Accommodating oneTBB (with hybrid cores support) that

myshevts
authored and
dmitrygo
committed
NCHW pooling perfomance fixed in accordance with v0.21

maxnick
authored and
dmitrygo
committed
[WA] Fixed fallback on ref conv in case exceeding scratchpad limit
dmitrygo
committed
Returned old behavior for fp32 avx2 1x1 conv with dw conv fusing

antonvor
authored and
dmitrygo
committed
Updated SoftPlus

a-sidorova
authored and
dmitrygo
committed
Fixed warning fo undefined ITT_ARCH_IA64 (#52)

ilya-lavrenov
authored and
dmitrygo
committed
Disable reorder JIT if both inputs and outputs are batch-strided.

IvanNovoselov
authored and
dmitrygo
committed
Include TBB headers as system

AlexPeskov
authored and
dmitrygo
committed
Fixed redifinition of tls model
dmitrygo
committed
nspc layout support for convolutions

maxnick
authored and
dmitrygo
committed
set scale = 1.f in case of signed input on platforms without vnni

antonvor
authored and
dmitrygo
committed
Enable direct copy primitives for u8 reorder

IvanNovoselov
authored and
dmitrygo
committed
Memory descriptor dynamism related changes

maxnick
authored and
dmitrygo
committed
Added prelu as binary post op

antonvor
authored and
dmitrygo
committed
Depthwise and Quantization post ops for Gemm Convolutions

antonvor
authored and
dmitrygo
committed
Perf fixes for Ref and NCHW Poolings

antonvor
authored and
dmitrygo
committed
perf fixes for quantization post ops

antonvor
authored and
dmitrygo
committed
todo: fix assert(idx < max_idx)

antonvor
authored and
dmitrygo
committed
simple reorder: temporarily disabled zero padding

antonvor
authored and
dmitrygo
committed
Fix possible data race when accessing global reorder list

Egor Duplensky
authored and
dmitrygo
committed
Brgemm implementation has perf degradation in RNN node

alexey-varyzgin
authored and
dmitrygo
committed
reverted old behavior with pdim_consistent check due to perf problems

antonvor
authored and
dmitrygo
committed
Renamed matmul kernel type: brg -> brgemm
dmitrygo
committed
Fixed bias addition order in brgemm kernel
dmitrygo
committed
[1D] Enlarge support

alexey-varyzgin
authored and
dmitrygo
committed
Update uni_ jit methods to avoid mixing vex and nonVEX instructions

Egor Duplensky
authored and
dmitrygo
committed
Quantization post op structure modified to reduce its complexity

maxnick
authored and
dmitrygo
committed
Hash utility functions were extracted to a separate module for reuse

maxnick
authored and
dmitrygo
committed
Desc similar_to routine consider start stride

maxnick
authored and
dmitrygo
committed
Desc similar_to routine use stride cmp mask

maxnick
authored and
dmitrygo
committed
added some legacy parallel methods to fix perf issues

antonvor
authored and
dmitrygo
committed
Migrate legacy post ops and zero points on runtime data pointers
dmitrygo
committed

Commits on Feb 21, 2022

Revert "all: remove mkldnn compatibility layer"
dmitrygo
committed

Commits on Mar 5, 2022

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

Commits on Mar 14, 2022

Fixed ODR violataion
dmitrygo
committed

Commits on Apr 5, 2022

equality_uni_xxx_for_sse_and_avx

chenhu-wang
authored and
dmitrygo
committed