Fix xpu2 kp compile error #53548
Fix xpu2 kp compile error #53548
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
| const uint32_t numel, | ||
| int read_lens) { | ||
| using Type = std::tuple_element_t<Index, ArgsT>; | ||
| #ifdef PADDLE_WITH_XPU_KP |
There was a problem hiding this comment.
89行的vec_size只与out有关吗？ 看你修改前的代码是与in/out同时有关的，不确定这里会不会隐藏性能问题
There was a problem hiding this comment.
参考elementwise也是只取了out的vec_size
There was a problem hiding this comment.
elementwise是因为dim是相同的，而broadcast 输入输出的dim可能是不同的……
There was a problem hiding this comment.
vec_size 原本是取 min( in out 4), 现在是取min( out 4)，那应该值是>=之前的值，所以应该不会造成性能下降，有其他原因考虑需要加上吗
| __simd__ ArgsT args[VecSize]; | ||
| __simd__ ConditionalT<OutT, NumOuts> result[VecSize]; | ||
|
|
||
| #ifdef PADDLE_WITH_XPU_KP |
There was a problem hiding this comment.
XPUKP 在broadcast的功能与GPU是一样的呀
There was a problem hiding this comment.
这里之前铭书对GPU的broadcast进行了特化优化（减少了其中重复的fast_divmod计算），这里为了保持其优化效果，就需要单独拿出来
| } else { | ||
| BcUnroller<BroadcastDataLoader, IsBoundary, LoadType, VecSize, Arity>::step( | ||
| ins, args, configs, use_broadcast, block_offset, num, numel); | ||
| ins, args, configs, use_broadcast, block_offset, num, numel, read_lens); |
There was a problem hiding this comment.
read_lens是给XPU KP 使用的，此处代码已经被else包含为什么还要添加read_lens
There was a problem hiding this comment.
上一个 comment 中，gpu部分做了特化，但 kp和 gpu 使用的是相同的非特化函数，参数就需要保持一致了
| } | ||
| #pragma unroll | ||
| for (int idx = 0; idx < read_lens; ++idx) { | ||
| std::get<Index>(dst[idx]) = in_temp[0]; |
There was a problem hiding this comment.
这里错误！ read_lens表示向量化读取的数量的格式，说明最终往dst里面存的数据应给是read_len个，而此处的只是循环写了in_temp 错误！！！
There was a problem hiding this comment.
这里没有 is boundary 的判断条件，不确定 read_lens 和Nx的关系，不知道具体是怎么读取数据，现在清楚已经修改
…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
3 participants
PR types
Bug fixes
PR changes
OPs
Description
fix compile error in xpu2