improve the performence of divide_double_grad #62533
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
你的PR提交成功，感谢你对开源项目的贡献! |
HydrogenSulfate
requested changes
Mar 8, 2024
| dy->Resize(y.dims()); | ||
| dev_ctx.template Alloc<T>(dy); | ||
| if (ddx_tensor == nullptr && ddy_tensor == nullptr) { | ||
| dy = nullptr; |
There was a problem hiding this comment.
- 如果推导右侧公式存在None导致左侧变量无法计算，根据现在paddle的方案，应该用全0填充：
FullLikeKernel<T, Context>(dev_ctx, y, Scalar(0.0), y.dtype(), dy);，否则大量代码需要修改。 - 把指针本身设置为空的行为应该是没有意义的
其他地方问题类似
| @@ -312,17 +312,6 @@ void DivideDoubleGradKernel(const Context& dev_ctx, | |||
| if (ddy_tensor == nullptr) { | |||
| dout = nullptr; | |||
There was a problem hiding this comment.
- 让传值的指针等于 nullptr 这个操作，应该是没意义的，进而可以把这里的if-else判断优化下
- dy应该赋值为形状跟y一样的全0矩阵
Comment on lines
180
to
190
| DenseTensor dz_div_y; | ||
| dz_div_y.Resize(out.dims()); | ||
| if (!dx_tensor || dx_tensor->dims() != out.dims()) { | ||
| dev_ctx.template Alloc<T>(&dz_div_y); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, grad_out, y, &dz_div_y, axis); | ||
| dx_tensor = &dz_div_y; | ||
| } |
There was a problem hiding this comment.
由于dx（dz_div_y），只有在算dy和dout的时候才会使用到，所以我觉得可以：
- 182行的if条件改为：
if ((dy || dout) && (!dx_tensor || dx_tensor->dims() != out.dims())) - 保持
dz_div_y的定义位置不变，将dz_div_y.Resize(out.dims());放到if条件里面，因为只有需要计算dx_tensor时，才需要dz_div_y这个中间变量
这样尽量减少不必要的计算。
| dy->Resize(y.dims()); | ||
| dev_ctx.template Alloc<T>(dy); | ||
| if (!ddx_tensor && !ddy_tensor) { | ||
| FullLikeKernel<T, Context>(dev_ctx, y, Scalar(0.0), y.dtype(), dy); |
There was a problem hiding this comment.
构造函数的第三个参数Scalar(0.0)是否可以改为Scalar(static_cast<T>(0.0))，否则当T为complex的时候可能会有问题？
| if (!ddx_tensor && !ddy_tensor) { | ||
| FullLikeKernel<T, Context>(dev_ctx, y, Scalar(0.0), y.dtype(), dy); | ||
| } else { | ||
| DenseTensor tmp_dy = tmp; |
There was a problem hiding this comment.
209行感觉可以删掉吧，反正tmp_dy也是语义不明，还不如直接用tmp，删掉之后下面的tmp_dy全部改成tmp
Comment on lines
211
to
259
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, *dx_tensor, y, &tmp_dy, axis); | ||
| if (ddx_tensor && !ddy_tensor) { | ||
| // dy = -dX * ddX / Y | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, *ddx_tensor, tmp_dy, dy, axis); | ||
| auto& place = *dev_ctx.eigen_device(); | ||
| auto dy_result = phi::EigenVector<T>::Flatten(*dy); | ||
| dy_result.device(place) = static_cast<T>(-1) * dy_result; | ||
| } else if (!ddx_tensor && ddy_tensor) { | ||
| // dY = Out * dX * ddY / Y | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, *ddy_tensor, tmp_dy, &tmp_dy, axis); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, out, tmp_dy, dy, axis); | ||
| } else { | ||
| // dY = Out * dX * ddY / Y - dX * ddX / Y | ||
|
|
||
| // dY = Out * dX * ddY / Y - dX * ddX / Y | ||
| phi::funcs::ElemwiseGradCompute<Context, T, DivGradDX<T>, DivDoubleDY<T>>( | ||
| dev_ctx, | ||
| ddX_safe, | ||
| ddY_safe, | ||
| out, | ||
| dX_div_Y, | ||
| axis, | ||
| nullptr, | ||
| dy, | ||
| DivGradDX<T>(), | ||
| DivDoubleDY<T>()); | ||
| // NOTE(dengkaipeng): in the following ElemwiseGradCompute, for the | ||
| // first output tensor is nullptr, the branch to calculate first | ||
| // output tensor will not be activated, DivGradDx function will not | ||
| // be called and can be ignored, the first branch has little effect | ||
| // on running speed. | ||
| phi::funcs:: | ||
| ElemwiseGradCompute<Context, T, DivGradDX<T>, DivDoubleDY<T>>( | ||
| dev_ctx, | ||
| *ddx_tensor, | ||
| *ddy_tensor, | ||
| out, | ||
| tmp_dy, | ||
| axis, | ||
| nullptr, | ||
| dy, | ||
| DivGradDX<T>(), | ||
| DivDoubleDY<T>()); | ||
| } | ||
| } |
There was a problem hiding this comment.
这里的逻辑看起来是首先统一在 开头求出公共项：dx/y，然后根据不同的条件走不同的分支，有的分支内部调用不止一次DefaultElementwiseOperator，所以这里看一下是否可以全部统一使用ElemwiseGradCompute，if-else分支内也通过一次调用计算完毕，只不过需要根据不同的条件，编写不同的dy_op（如DivDoubleDY_Only_DDX, DivDoubleDY_Only_DDY）并传给ElemwiseGradCompute，这些不同的dy_op运算时真正使用到的参数也是不同的，不使用的参数可以随便传一个形状相同的并且能正常访问的DenseTensor占位即可。
Comment on lines
268
to
306
| } else if (ddx_tensor != nullptr && ddy_tensor == nullptr) { | ||
| // ddOut = ddX / Y | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, *ddx_tensor, y, ddout, axis); | ||
| } else if (!ddx_tensor && ddy_tensor) { | ||
| // ddOut = - Out * ddY / Y | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, out, *ddy_tensor, &tmp, axis); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, tmp, y, ddout, axis); | ||
| auto& place = *dev_ctx.eigen_device(); | ||
| auto ddout_result = phi::EigenVector<T>::Flatten(*ddout); | ||
| ddout_result.device(place) = static_cast<T>(-1) * ddout_result; | ||
| } else { | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, out, *ddy_tensor, &tmp, axis); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::SubtractFunctor<T>, | ||
| funcs::InverseSubtractFunctor<T>>( | ||
| dev_ctx, *ddx_tensor, tmp, &tmp, axis); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, tmp, y, ddout, axis); | ||
| } |
There was a problem hiding this comment.
同理，多次DefaultElementwiseOperator调用是否可以优化成一次调用
| FullLikeKernel<T, Context>(dev_ctx, y, Scalar(0.0), y.dtype(), dy); | ||
| } else { | ||
| DenseTensor tmp_dy = tmp; | ||
| // dX / Y |
There was a problem hiding this comment.
// dX / Y ==> // pre-compute 'dX / Y' into 'tmp' for 'ddout' and/or 'dy'
| auto& place = *dev_ctx.eigen_device(); | ||
| auto dout_result = phi::EigenVector<T>::Flatten(*dout); | ||
| dout_result.device(place) = static_cast<T>(-1) * dout_result; | ||
| } | ||
| } | ||
| } |
| @@ -166,33 +166,28 @@ template <typename T, typename Context> | |||
| void DivideDoubleGradKernel(const Context& dev_ctx, | |||
There was a problem hiding this comment.
DivDoubleDY函数内可以把dout提出来，减少乘法次数
Comment on lines
165
to
175
| template <typename T> | ||
| struct DivDoubleDY_Only_DDX { | ||
| HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return -x * dout; } | ||
| }; | ||
|
|
||
| template <typename T> | ||
| struct DivDoubleDY_Only_DDY { | ||
| HOSTDEVICE T operator()(T x, T y, T out, T dout) const { | ||
| return y * out * dout; | ||
| } | ||
| }; |
|
代码覆盖率挂了，可能需要加一些单测来覆盖被标记为红色的代码：https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/10325058/job/25667520 |
HydrogenSulfate
requested changes
Mar 19, 2024
Comment on lines
+336
to
+350
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, out, *ddy_tensor, &tmp, axis); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::SubtractFunctor<T>, | ||
| funcs::InverseSubtractFunctor<T>>( | ||
| dev_ctx, *ddx_tensor, tmp, &tmp, axis); | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, tmp, y, ddout, axis); |
Comment on lines
320
to
334
| // ddOut = - Out * ddY / Y | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, out, *ddy_tensor, &tmp, axis); | ||
| // VLOG(4) << "5"; | ||
| funcs::DefaultElementwiseOperator<Context, | ||
| T, | ||
| funcs::DivideFunctor<T>, | ||
| funcs::InverseDivideFunctor<T>>( | ||
| dev_ctx, tmp, y, ddout, axis); | ||
| auto& place = *dev_ctx.eigen_device(); | ||
| auto ddout_result = phi::EigenVector<T>::Flatten(*ddout); | ||
| ddout_result.device(place) = static_cast<T>(-1) * ddout_result; |
Comment on lines
177
to
194
| // template <typename T> | ||
| // struct DivDoubleDDOut { | ||
| // HOSTDEVICE T operator()(T x, T y, T out, T dout) const { | ||
| // return (x - out * y) / dout; | ||
| // } | ||
| // }; | ||
|
|
||
| // template <typename T> | ||
| // struct DivDoubleDDOut_Only_DDX { | ||
| // HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return x / dout; } | ||
| // }; | ||
|
|
||
| // template <typename T> | ||
| // struct DivDoubleDDOut_Only_DDY { | ||
| // HOSTDEVICE T operator()(T x, T y, T out, T dout) const { | ||
| // return -out * y / dout; | ||
| // } | ||
| // }; |
| funcs::MultiplyFunctor<T>, | ||
| funcs::InverseMultiplyFunctor<T>>( | ||
| dev_ctx, out, *ddy_tensor, &tmp, axis); | ||
| // VLOG(4) << "5"; |
28 tasks
HydrogenSulfate
requested changes
Mar 30, 2024
zyfncg
approved these changes
Apr 1, 2024
HydrogenSulfate
approved these changes
Apr 1, 2024
# for free
to join this conversation on GitHub.
Already have an account?
# to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR Category
Performance Optimization
PR Types
Performance
Description
优化divide_double_grad的大算子实现，优化前：
优化后