* discard riscv innerproduct weight
* drop riscv conv convdw weight
* drop riscv deconv deconvdw weight

* discard weight memory for x86 and vulkan
* drop arm innerproduct weight
* drop arm convolution weight
* drop arm convolutiondepthwise weight
* drop x86 vulkan deconvolution deconvolutiondepthwise weight
* drop arm deconvolution deconvolutiondepthwise weight
* arm neon assembly optimization for innerproduct pack4

* unified relu avx512

* unifed clip avx512

* unaryop avx512

* sigmoid avx512

* binaryop avx512

* padding convolution avx512

* convolutiondepthwise avx512

* innerproduct avx512

* reshape avx512

* slice avx512

* hardsigmoid hardswish avx512

* swish avx512

* pooling avx512

* crop avx512

* convolution sgemm pack16

* convolution 3x3 winograd pack16

* interp avx512

* convolution sgemm pack1to16

* convolution sgemm pack16to8

* convolution sgemm pack8to16

* convolution sgemm pack16to4

* fix vulkan permute pack8

* fix vulkan convolution gemm pack8to1

* x86 sse2 optimization for conv1x1 conv3x3 pack4 and general sgemm pack4/pack4to1

* x86 sse2 optimization for conv3x3s1 pack4to1 and general sgemm convolution pack4to1, use aligned load/store

* enforce explicit alignment

* added a number of optimized sse layers, specifically to increase performance of mobilenet style networks

* use general fmadd

* forceline x86 fmadd for better performance

* fix msvc compile warning

* simplify swish implementation

* Use activation layer for better performance

* Optimize x86 ConvolutionDepthWise activation

* add general fused activation
* add NCNN_FORCE_INLINE option

* Fix codepaths for fp16 weights when only AVX is enabled

* Disable opt overrides

* Update SDK url

* Update vulkan SDK download version

* Debugging risv pad

* apply code-format changes

* fix padding test

* fix mips slice test

* fix lrn test

* implement mish swish image shader, fix pooling adaptive image storage support, drop debug output

* update ci ubuntu 18.04

Co-authored-by: nihui <shuizhuyuanluo@126.com>

* quantize and dequantize tests

* unify activation and usability function

* drop NCNN_REQUANT cmake option, test dequantize requantize pack8, fix webassembly build

* benchmark use requantize int8 model

* SSE2: BatchNorm

* Fixed batch norm in AVX configuration

* Optimized register size switch

* Attempt to pass CI

* Attempt to pass CI

* Bias op

* Element wise ops

* Support packing on x86 by default

* Fixed macro range in bias

* Use aligned read for packed data

* Update testutil.h

* Update pooling_x86.cpp

* Support wasn SIMD

* Fix emscripten compiler flags

* fix build

* more ci fix

* concat x86 pack4

* flatten x86 pack4

* more x86 pack4

* ci pass

* fix

* enable sse2 mathfun

* enable --experimental-wasm-simd

Co-authored-by: nihui <shuizhuyuanluo@126.com>
Co-authored-by: nihuini <nihuini@tencent.com>

* runtime cpu dispatch

* force thread one

* disable openmp for coverage

* simplify test layer

* print NCNN_TARGET_ARCH

* less ci build variants

* weight fp16 storage option

* test convdw int8

* apple a12 a13

* ncnn_add_layer ncnn_add_shader cmake macro

* added fp16 weight storage version

* Small changes

* Fixed fp16 weight storage layers

* fix innerproduct

* fix loop error

* Fix windows build.

Disable fp 16 conversion when detecting int8 weights.

Implement requested changes.

* Restyled by clang-format

* Restyled by astyle

* Restyled by clang-format

* Restyled by astyle

* Update option.cpp

Set fp16 storage based on vulkan being used or not.

* fix innerproduct activation location and add 4 parallel channel version

* Restyled by clang-format

* Restyled by astyle

* Restyled by clang-format

* Restyled by astyle

* revert arm file

* implement requested changes

* Restyled by clang-format

* Restyled by astyle

* Restyled by clang-format

* Restyled by astyle

Co-authored-by: Restyled.io <commits@restyled.io>

* added avx implementations of FC and Max pool

* Specify AVX2

* Small fixes and using Fused avx activations

* fix type casting

* fixing some CI errors

* Fix code format

* fix pooling test

* remove vector typedef

* More compile fixes

* remove vector typedef

* set c++ version to 17

* Force c++ 17

* Fixing mathfun

* Try and workaround typedef issues

* typefix

* Remove typedef

* switch to static inline

* attempting to fix msvc bug

* Verified MSVX FIX

* Fixing clang build

* commit before switch

* More avx and packing implementation

* Fix ctest

* starting the depthwise pack 8 implementation

* Unrolled loop

* add depthwise pack 8 implementations

* Working 1x1 pack 8 implementation added

* revert incorrect changes

* added conact elempack 8

* more elempack enabled layers added and started on the conversion of the winograd pack4 conv 3x3

* Added code formatting

* fix styling

* Unroll loops

* unrolling loops

* Added more elempac layers for mobilenet v3

* revert commit

* fix code style

* remove arm neon references

* remove pack4 references

* More cleanup

* added packing avx code

* fixing linux build ctests

* remove usage of aligned loads

* More aligned mem ops removed

* Cleanup, revert some files and remove not working winograd and shufflechannel implementation

* add stackoverflow referal

* Fix windows build

* Restyled by clang-format

* Restyled by astyle

* Restyled by clang-format

* Restyled by astyle

* implement requested chaanges

* remove reshape

* revert arm file change

* Restyled by clang-format

* Restyled by astyle

* Restyled by clang-format

* Restyled by astyle

* fix unterminated directive

Co-authored-by: Restyled.io <commits@restyled.io>

* fuse mish and depthwise convolution

* darknet2ncnn: support enet-coco conversion.

* fix requant relu6 bug

* fix

* delete pipeline change in forward/forward_inplace avoid race in multithreading

* initial the ncnn post training quantization tools

* clear some comments of tools

* fix the Travis ci compiler error

* optimize the conv sgemm int8 on arm64-v8a platform

* optimize int8 arm64-v8a with sadalp ins

* merge requantize op  into latest conv layer

* merge requantize op into conv-int8 op

* update the mobilenet.param in the benchmark

* Update README.md

update Kirin970 and RK3399

* try to fix the travis build error

* add the armv7a conv3x3s1 implement without overflow,remove old codes

* fix the bug of conv3x3s2 packed int8

* new int8 implement,weight quant by perchanel,better accuracy~

* fix the bug of conv3x3s1 packed int8 neon

* add the naive c fp32 and int8 winograd F(2,3)

* add the neon intrinsic int8 winograd F(2,3)

* optimize the armv7a int8 winograd F(2,3) with neon assembly

* optimize the armv7a int8 winograd F(2,3) input transform with assembly.

* add the requantize layer and int8 relu implement.

* add graph optimize conv1x1s2 -> conv1x1s1,begin optimize int8 aarch64.

* fix int8 bugs

* add the c naive im2col with sgemm

* add aarch64 int8 winograd f23, conv3x3s2 naive implement

* add the int8 sgemm conv7x7s2 on x86/armv7a platform

* optimize the int8 sgemm by neon intrinsic and packed kernel

* optimize the int8 sgemm with packed data

* optimize the int8 sgemm with armv7a neon assembly

* add the int8 sgemm on arm64-v8a platform

* perpare to merge latest codes from master

* add the int8 param files

* In the Class Net,add the fuse_network method