Speed/sequence op1 #9217
Speed/sequence op1 #9217
Conversation
dzhwinter force-pushed the speed/sequence_op1 branch from 22f414c to 2524003 Compare 
| Please add benchmark details between there two versions. |
luotao1 left a comment
luotao1 left a comment There was a problem hiding this comment.
请问8倍的加速比是对GPU来说的吧，CPU上维持不变？
| if (i == index[tid]) { | ||
| in_grad[item_dim * i + tid] = out_grad[tid]; | ||
| } else { | ||
| in_grad[item_dim * i + tid] = static_cast<T>(0); |
There was a problem hiding this comment.
如果先都赋值为0，再根据if条件，对in_grad[item_dim * i + tid] = out_grad[tid]，还会更快一点么？LastPool和FirstPool类似。
There was a problem hiding this comment.
不会。因为先都赋值为0将会多一次cuda kernel call。减少kernel call 会大大加快运行速度。
| # return x, lod, out | ||
| | ||
| # def compute(self, x, lod, out): | ||
| # self.attrs = {'pooltype': "FIRST"} |
| self.attrs = {'pooltype': "SUM"} | ||
| for i in range(4): | ||
| sub_x = x[lod[0][i]:lod[0][i + 1], :] | ||
| out[i] = sub_x.sum(axis=0) |
There was a problem hiding this comment.
test_seq_pool.py单测只是换了一些单测的顺序么？
| T, MaxPoolGradFunctor<T>><<<grid, threads, 0, context.stream()>>>( | ||
| MaxPoolGradFunctor<T>(), out_grad.data<T>(), | ||
| lod.CUDAData(context.GetPlace()), lod.size(), item_dim, | ||
| in_grad->mutable_data<T>(context.GetPlace()), index->data<int>()); |
There was a problem hiding this comment.
lod.CUDAData(context.GetPlace())和in_grad->mutable_data<T>(context.GetPlace())等可以先用临时变量定义在if条件前面么，这样303-338行的代码可以简短点。141-178行类似。
There was a problem hiding this comment.
I don't think that is a good way to save lines of code. One rule in the Google style is that keep the declaration as close as it used place. If we forward declared with a temporary variable, user has to find it when he read the following code.
| LGTM @qingqing01 Do you have any suggestions? |
* commit '33b8b3d22034423455a493712955e419aac7b19b': (251 commits) Remove redundant commands in build.sh and build_doc.sh Add dependencies Move v2/api/fluid to fluid/api and Adjust doc build commands Plain LRN op throws an exception when is_test is set in backward pass fix compiler error of profiler_test in ONLY_CPU mode fix server shutdown Translation for Model Configuration (PaddlePaddle#9513) Fix data transform when inplace (PaddlePaddle#9450) refine parallel add FAQ (PaddlePaddle#9494) Fix dist error with lr decay layer (PaddlePaddle#9489) add prefetch_op (PaddlePaddle#9495) Fix some errors (PaddlePaddle#9403) hookup WITH_FLUID_ONLY in TeamCity build.sh (PaddlePaddle#9509) Fix the order of reads and write from buffered channel (PaddlePaddle#9423) change WITH_FLUID to WITH_FLUID_ONLY (PaddlePaddle#9427) fix block num Revert "make append activation in place by default (PaddlePaddle#9417)" Speed/sequence op1 (PaddlePaddle#9217) fix a compile error ...
PaddlePaddle#9217) * [Auto Parallel] fix bugs for split_batches_for_accumulation && fix bugs for enable_delay_scale_loss * add enable_delay_scale_loss flag for auto_parallel * fix ut * Update ci_case_auto.sh
3 participants
fix #9099
every minibatch sequence_pool and sequence_pool_grad operator have a ~8x time acceleration.
for example,
the sequence_pool op enhanced from 0.815583 -> 0.119373
the sequence_pool_grad enhanced from 0.579614 -> 0.0830757
before optimize
after optimize