nihui
/
ncnn
Not watched
1
0
Fork 0
Code
Releases 46 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
459 Commits
5 Branches
354 MB
2 Download
Tree: a862a3432d
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'a862a3432d'
${ noResults }
Commit Graph

274 Commits (a862a3432d62585bf185ecdfccbb7167b0d09f0c)

Author SHA1 Message Date
  nihui a862a3432d aarch64 assembly for conv3x3s1, about 18% speed gain 7 years ago
  BUG1989 5f1e678e7f delete out.fill operate in the armv7 int8 convdw 7 years ago
  BUG1989 9b8e968139 add armv7 int8 convdw3x3 and conv3x3s2 left4 case 7 years ago
  nihuini ef36d79b7e implement the missing dequantize image on armv7, prefer neon-optimized 3-dim dequantize, fix #547 8 years ago
  nihuini 19ad4cf284 fix build without neon 8 years ago
  nihuini 6e02923489 add some comment 8 years ago
  nihuini 23de61fd07 as we already have the int8_scale_term switch, do not have to rely on the actual scale value 8 years ago
  nihuini dda21b8ea9 convolution depthwise 3x3 stride 1 with unroll 8 on aarch64 8 years ago
  nihuini 9e491446ce NCNN_XADD returns int, though rarely used, fix #538 8 years ago
  nihuini 144f818f88 Mat fill int optimization 8 years ago
  nihuini 8526a69777 packed int8 convolution 3x3 stride 1 for armv7, 7%~25% faster than vanilla one, but God knows how hard I try :| 8 years ago
  nihuini a76a07eb3f fix null sub group elemsize/allocator in depthwise deconvolution, fix #539 8 years ago
  nihuini 6f1b0b0a61 quantized padding in convolution, use range sweets 8 years ago
  nihuini bf1c58be46 padding is elemsize aware, copy_make_border is now a padding wrapper 8 years ago
  nihuini d321d65fc8 Mat range for zero-copy slicing reference 8 years ago
  nihuini f78e8fd8e4 use max threads as cpu count on openmp capable runtime 8 years ago
  nihuini 2790f3c640 fix bus error for int8 storage 8 years ago
  BUG1989 1b0e33460d add armv7 int8 conv3x3s1,using vaddw to replace vadd and vmovl 8 years ago
  nihui 72411b7a6c restore the old conv3x3s2 as reference, fast dilation convolution fails on striding 8 years ago
  nihui 1f20eb4e8c pack weight and more unroll makes improvement, ~20% faster for conv3x3s2 8 years ago
  chensy 30cc738309 fix asm "invalid operand" error for target iOS armv7 on file dequantize_arm.cpp 8 years ago
  Diego Gomes 4d73407df8 fix gettid call for glibc 8 years ago
  Diego Gomes 534f38ed87 fix auxv read for elf64 8 years ago
  nihuini 2dbaf6f7b7 store int8 scale in binary 8 years ago
  nihui fe14037777 more sub op preload 8 years ago
  nihui 2fe7ada4d8 add arm int8 convolution stub, preload group op for x86 8 years ago
  nihui eac7c66a97 fix fp32 group convolution on x86 8 years ago
  nihui 5d04a3a45c layer holds bottom blob scale, depthwise convolution read group scales 8 years ago
  nihui 354b95256c bump param version, backward compatible 8 years ago
  nihuini 2bc504925e fix int8_scales from multiple blobs, fix #512 8 years ago
  nihuini da352916fe fix pd using flag condition 8 years ago
  nihuini 6b536701c3 sub-mat shall be allocator-aware 8 years ago
  nihuini e34aa7786a armv7 int8 quantize/dequantize and conv1x1s1 8 years ago
  nihuini 4be27a0a89 int8 inference on x86 8 years ago
  nihui a169cec363 core int8 inference, quantize and dequantize, net using flag, caffe2ncnn reads int8 scale table 8 years ago
  nihui b6b90c888f
high resolution timestamp on windows 8 years ago
  nihui af49e2cada
install allocator.h 8 years ago
  nihui 7e1f358084
fix build on msvc 8 years ago
  nihui 9706cd1447 implement ncnn blob/workspace allocator, fine-grained per-layer openmp threads control, fix #469 8 years ago
  nihui 5879cb4d15
sgemm outperform direct conv on large channel 8 years ago
  nihuini 4b8101e7fc Revert "optimize interleave section for load first, about 5%~10% speed gain" 8 years ago
  nihui 56a667472a
sgemm is always faster on common channel size 8 years ago
  nihui 1e4eaeeacd optimize interleave section for load first, about 5%~10% speed gain 8 years ago
  nihui 6895cbf810 single vldm is faster than two vld1 on armv7, and some pipeline optimize 8 years ago
  nihuini 05d7562a5d reorder kernel weight, pipeline friendly ;) 8 years ago
  nihui b8f4f024a4 implement reorg yolodetectionoutput layer from caffe-yolov2 8 years ago
  nihuini ee98817446 proper first row/col handling in resize family, fix #429 8 years ago
  nihuini 511baa6718 optional image pixel api, fix #434 8 years ago
  nihuini 2368d29a1e more explicit alignment on armv7 8 years ago
  nihuini d172a34329 direct assembly port, enable convolution 1x1 sgemm on armv7 8 years ago