LV.1

JueXiao

25积分0赞

2 帖子 3 回复 0 收藏

TA的动态

TA的帖子

TA的回复

MLU端掉用多次__bang_mlp报错 我的回复：#1HelloAI回复可否贴一下报错的源码？#include "mlu.h" __mlu_entry__ void CmlpFloatMCPipeline_mlu(float* out_data, int16_t* in_data, int8_t* filter_data,int* in_channel,int* in_height,int* in_width,int* out_channel) { int in_data_num=in_channel[0]*in_height[0]*in_width[0]; int filter_data_num= in_channel[0]*in_height[0]*in_width[0]*out_channel[0]; int out_data_num=out_channel[0]; int per_in_data_num=in_channel[0]*in_height[0]*in_width[0]*32; int per_out_data_num=64; int per_filter_data_num=in_channel[0]*64; int loops_nums=filter_data_num/per_filter_data_num; int remain_nums=(filter_data_num-per_filter_data_num*loops_nums)/4096; int POS=0; __nram__ float nram_out_data[2*64]; __nram__ int16_t nram_in_data[5000]; __wram__ int8_t wram_filter_data[4096*64]; __memcpy(nram_in_data,in_data,in_data_num*sizeof(int16_t),GDRAM2NRAM); //L0 __memcpy(wram_filter_data,filter_data,per_filter_data_num*sizeof(int8_t),GDRAM2WRAM); printf("第一个filter:%d\\n",wram_filter_data[0]); //C0 __bang_mlp(nram_out_data, nram_in_data, wram_filter_data, 64, per_filter_data_num, POS); printf("第一个out:%f\\n",nram_out_data[0]); //L1 __memcpy(wram_filter_data,filter_data+ per_filter_data_num, per_filter_data_num*sizeof(int8_t),GDRAM2WRAM); int loops; for(loops=2;loops<loops_nums;loops++){ //Si-2 __memcpy(out_data+(loops-2)* per_out_data_num,nram_out_data+((loops-2)%2)* per_out_data_num, per_out_data_num*sizeof(float),NRAM2GDRAM); //Ci-1 __bang_mlp(nram_out_data +((loops-1)%2)*per_out_data_num, nram_in_data, wram_filter_data , 64, per_filter_data_num, POS); //Li __memcpy(wram_filter_data,filter_data+(loops)* per_filter_data_num, per_filter_data_num*sizeof(int8_t),GDRAM2WRAM); } //Sn-1 __memcpy(out_data+(loops_nums-2)* per_out_data_num,nram_out_data+((loops_nums-2)%2)* per_out_data_num, per_out_data_num*sizeof(float),NRAM2GDRAM); //Cn __bang_mlp(nram_out_data +(loops_nums-1)%2*per_out_data_num, nram_in_data, wram_filter_data +((loops_nums-1)%2)* per_filter_data_num , per_filter_data_num, per_filter_data_num, POS); //Li //Sn __memcpy(out_data+(loops_nums-1)* per_out_data_num,nram_out_data+((loops_nums-1)%2)* per_out_data_num, per_out_data_num*sizeof(float),NRAM2GDRAM); } 0

MLU370— BANG 3.0 编程优化直播课【FAQ合集】 我的回复：#4含含回复如果数据量超过NRAM，可以把计算任务切片成多个计算任务进行分次运行，并不限制切分方式有没有切片的demo程序呀？ 0

MLU370— BANG 3.0 编程优化直播课【FAQ合集】 我的回复：请问bangc中数据量太大超过NRAM内存有什么方法可以切分数据再运算吗 0

MLU370— BANG 3.0 编程优化直播课【FAQ合集】 我的回复：BANGC可以指定运行的Core吗 0