search

jerry20091103 / 23th-project

0 stars 0 forks source link

R2+1D 一層(我們自己寫IP) #49

Closed plumy2828 closed 2 years ago

plumy2828 commented 2 years ago

https://cpp.hotexamples.com/site/file?hash=0xfd8f5591aa0878aceabe4061fe804976cbefb8d273c70757fd494b9d202d7675&fullName=polybenchGpu-master/polybenchCodesCudaOpenClHMPPOpenAcc/HMPP/threedconv.c&project=lnangong/polybenchGpu

fann9876543210 commented 2 years ago

https://github.com/madhur13490/PolybenchGPU/blob/master/OpenCL/3DCONV/3DConvolution.c

plumy2828 commented 2 years ago

1.確認3D padding跟stride維度 2.把目前幾個小layer合成一個基本block

fann9876543210 commented 2 years ago

R2Plus1dStem layer pass synthesis

synthesis report:

image
plumy2828 commented 2 years ago

解決simulation error，確認weight的size

plumy2828 commented 2 years ago

已跑完全部第一層，我們的答案與golden誤差在小數點後五位

fann9876543210 commented 2 years ago

update synthesis report

image
jerry20091103 commented 2 years ago

把Interface直接加到r2plus1d.cpp上面了，四個io都是用axi memroy-mapped 目前效能應該蠻差的，建議輸入要先讀進buffer裡面，它才可以合成出burst的東西。weight到時候也可以直接死在ip裡 不過只是測試正確性效能應該沒差啦，IO的優化之後再慢慢做