additional math operations #64

New Issue

sb10q · 2020-07-17T10:03:37+08:00

sb10q commented

2020-07-17 10:03:37 +08:00

https://github.com/m-labs/artiq/issues/1485

sb10q added the

priority:medium

label 2020-07-17 10:03:37 +08:00

dpn commented

2020-07-30 08:05:04 +08:00

Contract items:

3.2.1 natural logarithm (base e) of a floating point number
3.2.2 logarithm (base 10) of a floating point number
3.2.3 base e exponentiation: ex, where x is a floating point number
3.2.4 exponentiation: xy, where x is a floating point number, for both the case where y is an integer and the case where y is a floating point number
3.2.5 sine of a floating point number
3.2.6 cosine of a floating point number
3.2.7 tangent of a floating point number
3.2.8 arcsine of a floating point number
3.2.9 arccosine of a floating point number
3.2.10 arctangent of a floating point number
3.2.11 matrix addition for 1D or 2D matrices of the same dimensions
3.2.12 matrix subtraction for 1D or 2D matrices of the same dimensions
3.2.13 matrix multiplication of an N x M matrix with an M x P matrix to produce an N x P matrix, where N, M, and P are positive integers (can be 1). This covers the case of multiplication of matrices, vectors, and combinations of matrices and vectors.
3.2.14 multiplication of a 1D or 2D matrix by a scalar of the same type as the matrix
3.2.15 transpose of a 1D or 2D matrix
~~3.2.16 determinant of a square 2D matrix~~
3.2.17 perform any of the mathematical functions 3.2.1 – 3.2.10 on a 1D or 2D matrix of floating point numbers, element-wise. For operation 3.2.4, the matrix elements can be either the base (x) or the exponent (y) in the operation.
3.2.18 cast all elements of a matrix to floating point numbers
3.2.19 cast all elements of a matrix to integers

Decide on/expose user-facing API for casting.
Provide core-device libm functions necessary for the above.

Missing features and limitations not explicitly mentioned in #1485 that are probably just about necessary for real-world use:

Embedding of host values
RPCs
array() from other TArrays
Decide whether to always expand all iterables in array() (i.e. assume them to be rectangular)
Multi-dimensional indexing
core_log for multidimensional arrays (or just don't support)
Reshaping (as vec.T is the same 1D vec in NumPy)
~~Throw range/domain errors? (log, asin, ... just return NaN in NumPy, but check all)~~ They seem to just generate RuntimeWarnings and return NaN (the latter of which we do already).
Fix numpy.full

Nice-to-haves:

arctan2 (which is almost always what one really wants) and other binary libm functions
Expose Bessel functions/... available from libm from their respective SciPy symbols
Add more array constructors ({empty, zeros, ones}[_like])
Basic support for comparing arrays and associated bool array handling:
Array comparison (or cleanly forbid)
all/any
Negation of bool ararys

Tests to add:

Array shape mismatch in binops
Full feature coverage, comparing kernel/host
String element types (shouldn't be expanded)
Explicit dtypes
numpy.float()
Assigning binops

Contract items: - [x] 3.2.1 natural logarithm (base e) of a floating point number - [x] 3.2.2 logarithm (base 10) of a floating point number - [x] 3.2.3 base e exponentiation: ex, where x is a floating point number - [x] 3.2.4 exponentiation: xy, where x is a floating point number, for both the case where y is an integer and the case where y is a floating point number - [x] 3.2.5 sine of a floating point number - [x] 3.2.6 cosine of a floating point number - [x] 3.2.7 tangent of a floating point number - [x] 3.2.8 arcsine of a floating point number - [x] 3.2.9 arccosine of a floating point number - [x] 3.2.10 arctangent of a floating point number - [x] 3.2.11 matrix addition for 1D or 2D matrices of the same dimensions - [x] 3.2.12 matrix subtraction for 1D or 2D matrices of the same dimensions - [x] 3.2.13 matrix multiplication of an N x M matrix with an M x P matrix to produce an N x P matrix, where N, M, and P are positive integers (can be 1). This covers the case of multiplication of matrices, vectors, and combinations of matrices and vectors. - [x] 3.2.14 multiplication of a 1D or 2D matrix by a scalar of the same type as the matrix - [x] 3.2.15 transpose of a 1D or 2D matrix - <del>[ ] 3.2.16 determinant of a square 2D matrix</del> - [x] 3.2.17 perform any of the mathematical functions 3.2.1 – 3.2.10 on a 1D or 2D matrix of floating point numbers, element-wise. For operation 3.2.4, the matrix elements can be either the base (x) or the exponent (y) in the operation. - [x] 3.2.18 cast all elements of a matrix to floating point numbers - [x] 3.2.19 cast all elements of a matrix to integers --- - [x] Decide on/expose user-facing API for casting. - [x] Provide core-device libm functions necessary for the above. --- Missing features and limitations not explicitly mentioned in #1485 that are probably just about necessary for real-world use: - [x] Embedding of host values - [x] RPCs - [x] array() from other TArrays - [x] Decide whether to always expand all iterables in `array()` (i.e. assume them to be rectangular) - [x] Multi-dimensional indexing - [x] core_log for multidimensional arrays (or just don't support) - [ ] Reshaping (as vec.T is the same 1D vec in NumPy) - [x] <del>Throw range/domain errors? (log, asin, ... just return NaN in NumPy, but check all)</del> They seem to just generate RuntimeWarnings and return NaN (the latter of which we do already). - [x] Fix `numpy.full` --- Nice-to-haves: - [x] arctan2 (which is almost always what one really wants) and other binary libm functions - [x] Expose Bessel functions/... available from libm from their respective SciPy symbols - [ ] Add more array constructors (`{empty, zeros, ones}[_like]`) - Basic support for comparing arrays and associated bool array handling: - [ ] Array comparison (or cleanly forbid) - [ ] all/any - [ ] Negation of bool ararys --- Tests to add: - [ ] Array shape mismatch in binops - [x] Full feature coverage, comparing kernel/host - [x] String element types (shouldn't be expanded) - [ ] Explicit dtypes - [ ] `numpy.float()` - [x] Assigning binops

pca006132 commented

2020-08-26 14:32:05 +08:00

If we want to support faster math operations on slices, maybe we can have a look at NEON optimized math libraries like Ne10 later. SIMD really improves the performance by a lot if we need to perform operations on large arrays.

One problem with the Ne10 library is that it seems to be dead for 2 years. Other libraries seems to target AArch64 instead of Cortex-A devices.

If we want to support faster math operations on slices, maybe we can have a look at NEON optimized math libraries like [Ne10](https://github.com/projectNe10/Ne10) later. SIMD really improves the performance by a lot if we need to perform operations on large arrays. One problem with the Ne10 library is that it seems to be dead for 2 years. Other libraries seems to target AArch64 instead of Cortex-A devices.

sb10q commented

2020-08-26 19:13:29 +08:00

Poster

Other libraries seems to target AArch64 instead of Cortex-A devices.

If we have Zynq Ultrascale+ devices at some point, we could use those.

> Other libraries seems to target AArch64 instead of Cortex-A devices. If we have Zynq Ultrascale+ devices at some point, we could use those.

dpn commented

2020-08-26 21:19:15 +08:00

The LLVM autovectorizer should handle array operations quite nicely – we might need to enable a higher optimization level (and possibly switch to a newer LLVM), though.

dpn commented

2020-11-11 04:10:41 +08:00

core_log will simply log things row by row, i.e. as array([array([…]), …]), which is imho acceptable, given how rare that use case will actually be in practice. Just wanted to make sure it doesn't crash and burn. Also covered by the test suite already.

`core_log` will simply log things row by row, i.e. as `array([array([…]), …])`, which is imho acceptable, given how rare that use case will actually be in practice. Just wanted to make sure it doesn't crash and burn. Also covered by the test suite already.

sb10q commented

2021-01-19 14:48:11 +08:00

Poster

@dpn Is there still more to be done here or should we close this issue?