
Hi! Problem: Using Fortran routines from Python C/API is "tricky" when multi-dimensional arrays are passed in. Cause: Arrays in Fortran are stored in column-wise order while arrays in C are stored in row-wise order. Standard solutions: 1) Create a new C array; copy the data from the old one in column-wise order; pass the new array to fortran; copy changed array back to old one in row-wise order; deallocate the array. 2) Change the storage order of an array in place: element-wise swapping; pass the array to fortran; change the storage order back with element-wise swapping Why standard solutions are not good? 1) Additional memory allocation, that is problem for large arrays; Element-wise copying is time consuming (2 times). 2) It is good as no extra memory is needed but element-wise swapping (2 times) is approx. equivalent with the element-wise copying (4 times). Proclamation: Introduce a column-wise array to Numeric Python where data is stored in column-wise order that can be used specifically for fortran routines. Proposal sketch: 1) Introduce a new flag `row_order'to PyArrayObject structure: row_order == 1 -> the data is stored in row-wise order (default, as it is now) row_order == 0 -> the data is stored in column-wise order Note that now the concept of contiguousness depends on this flag. 2) Introduce new array "constructors" such as PyArray_CW_FromDims, PyArray_CW_FromDimsAndData, PyArray_CW_ContiguousFromObject, PyArray_CW_CopyFromObject, PyArray_CW_FromObject, etc. that all return arrays with row_order=0 and data stored in column-wise order (that is in case of contiguous results, otherwise strides feature is employd). 3) In order to operations between arrays (possibly with different storage order) would work correctly, many internal functions of NumPy C/API need to be modifyied. 4) anything else? What is the good of this? 1) The fact is that there is a large number of very good scietific tools freely available written in Fortran (Netlib, for instance). And I don't mean only Fortran 77 codes but also Fortran 90/95 codes. 2) Having Numeric Python arrays with data stored in column-wise order, calling Fortran routines from Python becomes really efficient and space-saving. 3) There should be little performance hit if, say, two arrays with different storage order are multiplied (compared to the operations between non-contiguous arrays in the current implementation). 4) I don't see any reason why older C/API modules would broke because of this change if it is carried out carefully enough. So, back-ward compability should be there. 5) anything else? What are against of this? 1) Lots of work but with current experience it should not be a problem. 2) The size of the code will grow. 3) I suppose that most people using Numerical Python will not care of calling Fortran routines from Python. Possible reasons: too "tricky" or no need. In the first case, the answer is that there are tools such as PyFort, f2py that solve this problem. In the later case, there is no problem:-) 4) anything else? I understand that my proposal is quite radical but taking into account that we want to use Python for many years to come, the use would be more pleasing if one cause of (constant) confusion would be less during this time. Best regards, Pearu

This is a very interesting proposal that we should consider carefully. I seem to recall reading that Jim Hugunin originally had this idea in mind when he established the concept of contiguousness, etc. My current thoughts on this issue are that it is of only syntatic value and seems like a lot of extra code has to be written in order to provide this "user-friendliness." I don't see why it is so confusing to recognize that Fortran just references it's arrays "backwards" (or Python references them backwards --- whatever your preference). How you index into an array is an arbitrary decision. Numerical Python and Fortran just have opposite conventions. As long as that is clear, I don't see the real trouble. If the Fortran documentation calls for an array of dimension (M,N,L) you pass it a contiguous Python array of shape (L,N,M) --- pretty simple. Perhaps someone could enlighten me as to why this is more than just a aesthetic problem. Right now, I would prefer that the time spent by someone to "fix" this "problem" went to expanding the availability of easy-to-use processing routines for Numerical Python, or improving the cross-platform plotting capabilities. I agree that it can be most confusing when you are talking about matrix math since we are so used to thinking of matrix multiplication as A * B = C with a shape analysis of: M X N * N X L = M X L If the matrix multiplacation code is in Fortran, then it expects to get an (M,N) array and a (N,L) array and returns an (M,L) array. But from Python you would pass it arrays with shape (N,M) and (L,N) and get back an (L,M) array which can be confusing to our "understanding" of shape analysis in matrix multiplication: Python matrix multiplication rule if calling a Fortran routine to do the multiplication: (N,M) (L,N) = (L, M) I think a Python-only class could solve this problem much more easily than changing the underlying C-code. This new Python Fortran-array class would just make the user think that the shapes were (M,N) and (N,L) and the output shape was (M,L). For future reference, any array-processing codes that somebody writes should take a strides array as an argument, so that it doesn't matter what "order" the array is in. --Travis

On Wed, 26 Jan 2000, Travis Oliphant wrote:
I think that this expansion would be quicker if the Python/Fortran connection would not introduce this additional question to worry about.
I can see the following problems when two different conventions are mixed: 1) if your application Python code is larger than "just an example that demonstrates the correct usage of two different conventions" and it can call other C/API modules that do calculations in C convention then you need some kind of book keeping where your matrices need to be transposed and where not, and where to insert additional code for doing transposition. I think this can be done in lower level and more efficiently than most ordinary users would do anyway. 2) Another but minor drawback of having two conventions is that if you have square matrix that is non-symmetric, then its misuse would be easy and (may be) difficult to discover. On the other hand, I completely understand why my proposal would not be implemented --- it looks like it needs lots of work and in short term the gain would not be visible to most users. Pearu

This is a very interesting proposal that we should consider carefully. I seem to recall reading that Jim Hugunin originally had this idea in mind when he established the concept of contiguousness, etc. My current thoughts on this issue are that it is of only syntatic value and seems like a lot of extra code has to be written in order to provide this "user-friendliness." I don't see why it is so confusing to recognize that Fortran just references it's arrays "backwards" (or Python references them backwards --- whatever your preference). How you index into an array is an arbitrary decision. Numerical Python and Fortran just have opposite conventions. As long as that is clear, I don't see the real trouble. If the Fortran documentation calls for an array of dimension (M,N,L) you pass it a contiguous Python array of shape (L,N,M) --- pretty simple. Perhaps someone could enlighten me as to why this is more than just a aesthetic problem. Right now, I would prefer that the time spent by someone to "fix" this "problem" went to expanding the availability of easy-to-use processing routines for Numerical Python, or improving the cross-platform plotting capabilities. I agree that it can be most confusing when you are talking about matrix math since we are so used to thinking of matrix multiplication as A * B = C with a shape analysis of: M X N * N X L = M X L If the matrix multiplacation code is in Fortran, then it expects to get an (M,N) array and a (N,L) array and returns an (M,L) array. But from Python you would pass it arrays with shape (N,M) and (L,N) and get back an (L,M) array which can be confusing to our "understanding" of shape analysis in matrix multiplication: Python matrix multiplication rule if calling a Fortran routine to do the multiplication: (N,M) (L,N) = (L, M) I think a Python-only class could solve this problem much more easily than changing the underlying C-code. This new Python Fortran-array class would just make the user think that the shapes were (M,N) and (N,L) and the output shape was (M,L). For future reference, any array-processing codes that somebody writes should take a strides array as an argument, so that it doesn't matter what "order" the array is in. --Travis

On Wed, 26 Jan 2000, Travis Oliphant wrote:
I think that this expansion would be quicker if the Python/Fortran connection would not introduce this additional question to worry about.
I can see the following problems when two different conventions are mixed: 1) if your application Python code is larger than "just an example that demonstrates the correct usage of two different conventions" and it can call other C/API modules that do calculations in C convention then you need some kind of book keeping where your matrices need to be transposed and where not, and where to insert additional code for doing transposition. I think this can be done in lower level and more efficiently than most ordinary users would do anyway. 2) Another but minor drawback of having two conventions is that if you have square matrix that is non-symmetric, then its misuse would be easy and (may be) difficult to discover. On the other hand, I completely understand why my proposal would not be implemented --- it looks like it needs lots of work and in short term the gain would not be visible to most users. Pearu
participants (2)
-
Pearu Peterson
-
Travis Oliphant