CPU forward calculation replaces Eigen with Lapack;Modify linalg exposure rules #35916
Conversation
|
Thanks for your contribution! |
Zjq9409
changed the title
| input_tensor = dito.Transpose(input); | ||
| math::DeviceIndependenceTensorOperations<platform::CPUDeviceContext, T>( | ||
| ctx); | ||
| *eigen_vectors = dito.Transpose(input); |
There was a problem hiding this comment.
- lapack输入是列优先的？最好加一些注释说明，这里为什么要transpose。
- 看了下svd_helper.h里面实现的Transpose，既然只针对最好2维进行transpose，其实没必要区分2-6 D，eigvals_op.h里面
SpiltBatchSquareMatrix这种实现方式比较好。 EigvalsKernel里面，math::lapackEig<T, Real<T>>这种调用方式也比较好，能够清楚的体现values、vectors的数据类型的关系。
There was a problem hiding this comment.
如果has_vectors为false，好像是否转置得到的特征值都是一样的，所以这里可能不需要transpose？
| template <typename ValueType, typename T> | ||
| struct MatrixEighFunctor<platform::CPUDeviceContext, ValueType, T> { | ||
| public: | ||
| void operator()(const framework::ExecutionContext &ctx, const Tensor &input, |
There was a problem hiding this comment.
functor里面不再需要从ctx里面获取运行时信息，最好传DeviceContext。
| int liwork = -1; | ||
| int iwork_opt = -1; | ||
| T lwork_opt = static_cast<T>(-1); | ||
| ValueType rwork_opt = static_cast<ValueType>(-1); |
There was a problem hiding this comment.
这些参数都是啥意思？-1代表什么？lapack函数中会写这些参数吗？可以加一些注释说明下。
| auto *infos_data = info_tensor.mutable_data<int>( | ||
| framework::make_ddim({batch_size}), ctx.GetPlace()); | ||
|
|
||
| math::lapackEvd<T, ValueType>(jobz, uplo, n, out_vector, lda, out_value, |
There was a problem hiding this comment.
加个注释说明单独这一次lapack函数调用的目的是什么。
| *info_ptr)); | ||
| } | ||
| if (has_vectors) { | ||
| *eigen_vectors = dito.Transpose(*eigen_vectors); |
| auto *value_data = out_value + i * values_stride; | ||
| auto *vector_data = out_vector + i * vector_stride; | ||
| int *info_ptr = &infos_data[i]; | ||
| math::lapackEvd<T, ValueType>(jobz, uplo, n, vector_data, lda, value_data, |
There was a problem hiding this comment.
Evd是什么意思，会被用到别的计算吗，是否直接叫lapackEigh比较好？
| T lwork_opt = static_cast<T>(-1); | ||
| ValueType rwork_opt = static_cast<ValueType>(-1); | ||
|
|
||
| Tensor info_tensor; |
There was a problem hiding this comment.
如果一次只计算1个矩阵，貌似不用定义一个info_tensor，而是直接用int info就够？
| PADDLE_ENFORCE_EQ( | ||
| *info_ptr, 0, | ||
| platform::errors::PreconditionNotMet( | ||
| "For batch [%d]: the [%d] argument had an illegal value", i, |
| } | ||
|
|
||
| template <> | ||
| void lapackEvd<paddle::platform::complex<float>, float>( |
Zjq9409
force-pushed
the
LapackEigh
branch
2 times, most recently
from
d4df8ec to
a0c1b49
Compare
There was a problem hiding this comment.
has_vectors为false时，eigen_vectors为nullptr，不应该调mutable_data。
| input_tensor = dito.Transpose(input); | ||
| math::DeviceIndependenceTensorOperations<platform::CPUDeviceContext, T>( | ||
| ctx); | ||
| *eigen_vectors = dito.Transpose(input); |
There was a problem hiding this comment.
如果has_vectors为false，好像是否转置得到的特征值都是一样的，所以这里可能不需要transpose？
There was a problem hiding this comment.
可以using ValueType = math::Real<T>;，这样模板里面不用传ValueType。
3 participants
PR types
Others
PR changes
OPs
Describe
Eigh算子前向计算将eigen库替换为Lapack库实现，修改linalg暴露规则。