Tencent / ncnn

ncnn is a high-performance neural network inference framework optimized for the mobile platform
Other
19.8k stars 4.11k forks source link

Conv层结果不一致 #5479

Open XdpAreKid opened 1 month ago

XdpAreKid commented 1 month ago

error log | 日志或报错信息 | ログ

value not match at c:0 d:0 h:31 w:24 expect 2.745680 but got 2.427734 test_layer_gpu failed test_layer Convolution failed use_packing_layout=1 use_fp16_packed=1 use_fp16_storage=0 use_fp16_arithmetic=0 use_shader_pack8=0 use_bf16_storage=1 use_image_storage=0 use_sgemm_convolution=1 use_winograd_convolution=1 test_convolution failed w=34 h=34 c=1024 outch=1024 kernel=3 dilation=1 stride=1 pad=0 bias=1 act=5 actparams=[-0.719921,0.291636]

context | 编译/运行环境 | バックグラウンド

how to reproduce | 复现步骤 | 再現方法

1.增加test_convolution(34, 34, 1024, 1024, 3, 1, 1, 0, 1); 2. 3.

more | 其他 | その他

nihui commented 1 month ago

什么系统,硬件?cpu/gpu型号? 是 20240410 版本?

XdpAreKid commented 1 month ago

什么系统,硬件?cpu/gpu型号? 是 20240410 版本?

archlinux, i9-9900K+RTX 3060 , master和20240102均存在