dropout_nd_optimization #51479
dropout_nd_optimization #51479
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
zhangboSJTU
force-pushed
the
dropout_optimize
branch
from
84f61c9 to
a66a971
Compare
| T dst_mask[kCount]; // 0 ~ kCount - 1 : dst, kCount ~ 2 * kCount - 1: mask | ||
| float rands[kCount]; | ||
| MaskType mask_result[kCount]; | ||
| uint8_t mask_result[kCount]; |
There was a problem hiding this comment.
MaskType 已经作为 uint8_t ，由模板传入了，这里不需要替换成uint8_t.
| uint32_t offset = 0u; | ||
| uint32_t idx = i; | ||
| // Use (j < phi::DDim::kMaxRank) conditiion rather than | ||
| // (j < broadcast_config.rank) for (#pragma unroll) |
There was a problem hiding this comment.
#pragma unroll 的位置放错了，应该紧贴 for-loop
There was a problem hiding this comment.
后面删除注释可能不小心调整位置了，根据建议修改
| template <typename T> | ||
| struct DstFunctor { | ||
| using MT = typename phi::kps::details::MPTypeTrait<T>::Type; | ||
| MT factor; |
There was a problem hiding this comment.
factor 调整为 private 变量，如果HOSTDEVICE inline DstFunctor(const float retain_prob, 是在Host上执行的话，不需要HOSTDEVICE inline
There was a problem hiding this comment.
DstFunctor是已有的代码，没有任何需要调整的场景下，不要挪动位置
There was a problem hiding this comment.
之前代码functor和函数排列太混乱，删除无用函数后调整了下
There was a problem hiding this comment.
factor调整为 private 变量，如果HOSTDEVICE inline DstFunctor(const float retain_prob,是在Host上执行的话，不需要HOSTDEVICE inline
done
| } | ||
| } | ||
| }; | ||
|
|
There was a problem hiding this comment.
同上，已有的代码，如果不涉及大面积修改，不要移动
| if (rand[i] < retain_prob_) { | ||
| dst[i] = static_cast<T>(1); | ||
| } else { | ||
| dst[i] = static_cast<T>(0); |
There was a problem hiding this comment.
dst[i] = (rand[i] < retain_prob_) ? static_cast<T>(1) : static_cast<T>(0);
| } | ||
| }; | ||
|
|
||
| template <typename T> |
There was a problem hiding this comment.
为什么取消了
template <typename T1, typename T2 = T1, typename OutT = T1>
There was a problem hiding this comment.
该函数只在此处使用，没必要加过多无用模版
There was a problem hiding this comment.
原始代码中存在kps::OperatorTernary<T, float, T, DstMaskFunctor<T, float>> ，感觉又是一个无必要修改的地方
| }; | ||
|
|
||
| template <typename T, typename MaskType> | ||
| template <typename T> |
There was a problem hiding this comment.
typename MaskType保留，显示的uint8_t 代码后续失去了快速扩展性.
There was a problem hiding this comment.
考虑到这个函数的功能特定，优化后又做了kernel融合，不需要过多的模版参数
|
PR-CI-ROCM-Compile 我rerun了两轮还是有问题，找个设备检查下 |
报错同 |
zhangboSJTU
force-pushed
the
dropout_optimize
branch
from
7129e50 to
a66a971
Compare
* with printf * add DropOutNdForwardKernel * PR comment
…to Release/2.5 (#53623) * Support different dtypes of inputs for broadcast for dropout optimization (#52093) * change judgement for DropoutGradGPUKernelDriver * add UnrollerWithoutVecSize and after this Loaddata to be refined * pass unittest * use same unroller with XPU * BroadcastWithInt64Index * BroadcastDataLoader template partial specialization * fix compile errs in ROCms * PR comment * dropout_nd_optimization (#51479) * with printf * add DropOutNdForwardKernel * PR comment * Dropout optimize & clean broadcast inT and ElementwiseType (#52969) * change judgement for DropoutGradGPUKernelDriver * add UnrollerWithoutVecSize and after this Loaddata to be refined * pass unittest * use same unroller with XPU * BroadcastWithInt64Index * BroadcastDataLoader template partial specialization * fix compile errs in ROCms * clean ElementwiseT and InT for BroadcastKernel * default axis and clean inT * remove redundant fast divmod computation * optimize drop_nd & drop_nd_grad * optimize BroadcastDataLoader bf16 fp16 * rm InT etc. after merge develop * delete constexpr for windows ci * fix conflict * fix conflic with develop * fix conflic * new clean * clean * Fix xpu2 kp compile error (#53548) * fix conflict * conflict
2 participants
PR types
Performance optimization
PR changes
OPs
Describe
Dropout_nd performance optimization