pystencils merge requestshttps://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests2019-09-26T17:12:44+02:00https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/60Eliminate usages of old name 'equation collection' for `AssignmentCollection`2019-09-26T17:12:44+02:00Stephan SeitzEliminate usages of old name 'equation collection' for `AssignmentCollection`We should avoid the old name equation collection.We should avoid the old name equation collection.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/59Document backends.json2019-09-26T12:49:19+02:00Stephan SeitzDocument backends.jsonhttps://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/57Add AssignmentCollection.{free_fields,bound_fields}2019-09-25T15:41:44+02:00Stephan SeitzAdd AssignmentCollection.{free_fields,bound_fields}Wasn't this merged already?Wasn't this merged already?https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/56Interpolation 24.0.92019-09-25T15:41:24+02:00Stephan SeitzInterpolation 24.0.9This is another rebased PR for integrating interpolated accesses.
Iterpolation accesses work like `absolute_access` except they can be savely applied on all fields (i.e. with boundary checks).
More info here: !20
This PR contains som...This is another rebased PR for integrating interpolated accesses.
Iterpolation accesses work like `absolute_access` except they can be savely applied on all fields (i.e. with boundary checks).
More info here: !20
This PR contains some dead code that uses https://github.com/theHamsta/CubicInterpolationCUDA . I have not included it as a submodule in pystencils in this PR.
This PR break the hash of those two test:
```
[gw11] [ 14%] FAILED lbmpy_tests/test_code_hashequivalence.py::test_hash_equivalence_llvm
lbmpy_tests/test_conserved_quantity_relaxation_invariance.py::test_srt
[gw8] [ 15%] FAILED lbmpy_tests/test_code_hashequivalence.py::test_hash_equivalence
```https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/58Extra asserts sympy issue2019-09-25T15:38:17+02:00Stephan SeitzExtra asserts sympy issueAdd extra assertions to be super sure.Add extra assertions to be super sure.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/20WIP: Astnodes for interpolation2019-09-24T18:59:50+02:00Stephan SeitzWIP: Astnodes for interpolationThis PR needs maybe still needs some clean-up.
However, it would be good to recieve already some feed-back.
What works:
- Using CUDA textures
- Using HW accelerated interpolation for float32 textures
- Implement linear interpol...This PR needs maybe still needs some clean-up.
However, it would be good to recieve already some feed-back.
What works:
- Using CUDA textures
- Using HW accelerated interpolation for float32 textures
- Implement linear interpolations either via software (CPU, GPU), texture accesses without HW-interpolation but HW boundary handling
- Adding transformed coordinate systems to fields
What does not work:
- HW boundary handling for CUDA textures for the boundary handling modes `mirror` and `wrap` (apparently they have been removed from CUDA's API but are still present in pycuda. Now there's only
```
cudaBoundaryModeZero = 0
Zero boundary mode
cudaBoundaryModeClamp = 1
Clamp boundary mode
cudaBoundaryModeTrap = 2
Trap boundary mode
```
Wtf is trap boundary mode? Nothing is documented so we can only experiment.
What kind of works:
- B-Spline interpolation on GPU using this repo as a submodule (http://www.dannyruijters.nl/cubicinterpolation/), to lazy for tests. Don't know how to prove correctness
- Textures for dtypes with itemsize > 4. PyCUDA has helper header (https://github.com/inducer/pycuda/blob/master/pycuda/cuda/pycuda-helpers.hpp) that loads doubles by two int fetches. However, this hack seems to be only working if we add a 0.5 offset and make all functions in this header accept float.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/43Use get_type_of_expression in typing_form_sympy_inspection to infer types2019-09-23T16:16:50+02:00Stephan SeitzUse get_type_of_expression in typing_form_sympy_inspection to infer typesDANGER ZONE: this changes something in the core behavior of pystencils. Be careful before merging!
In summary, when `typing_form_sympy_inspection` reaches the point where it would just use `default_type`, we try to use `get_type_of_ex...DANGER ZONE: this changes something in the core behavior of pystencils. Be careful before merging!
In summary, when `typing_form_sympy_inspection` reaches the point where it would just use `default_type`, we try to use `get_type_of_expression` to infer the actual type.
We use information of previously defined variables in current scope.
Another approach would be to just type all the intermediate variable with `auto`.
```python
x = pystencils.fields('x: float32[3d]')
assignments = pystencils.AssignmentCollection({
a: cast_func(10, create_type('float64')),
b: cast_func(10, create_type('uint16')),
e: 11,
c: b,
f: c + b,
d: c + b + x.center + e,
x.center: c + b + x.center
})
```
Before:
```cpp
FUNC_PREFIX void kernel(float * RESTRICT _data_x, int64_t const _size_x_0, int64_t const _size_x_1,
int64_t const _size_x_2, int64_t const _stride_x_0, int64_t const _stride_x_1, int64_t const _stri
de_x_2)
{
const double a = 10.0;
const double b = 10;
const double e = 11.0;
const double c = b;
const double f = b + c;
for (int ctr_0 = 0; ctr_0 < _size_x_0; ctr_0 += 1)
{
float * RESTRICT _data_x_00 = _data_x + _stride_x_0*ctr_0;
for (int ctr_1 = 0; ctr_1 < _size_x_1; ctr_1 += 1)
{
float * RESTRICT _data_x_00_10 = _stride_x_1*ctr_1 + _data_x_00;
for (int ctr_2 = 0; ctr_2 < _size_x_2; ctr_2 += 1)
{
const double d = b + c + e + _data_x_00_10[_stride_x_2*ctr_2];
_data_x_00_10[_stride_x_2*ctr_2] = b + c + _data_x_00_10[_stride_x_2*ctr_2];
}
}
}
}
```
After:
```cpp
FUNC_PREFIX void kernel(float * RESTRICT _data_x, int64_t const _size_x_0, int64_t const _size_x_1,
int64_t const _size_x_2, int64_t const _stride_x_0, int64_t const _stride_x_1, int64_t const _stri
de_x_2)
{
const double a = 10.0;
const uint16_t b = 10;
const int64_t e = 11.0;
const uint16_t c = b;
const uint16_t f = b + c;
for (int ctr_0 = 0; ctr_0 < _size_x_0; ctr_0 += 1)
{
float * RESTRICT _data_x_00 = _data_x + _stride_x_0*ctr_0;
for (int ctr_1 = 0; ctr_1 < _size_x_1; ctr_1 += 1)
{
float * RESTRICT _data_x_00_10 = _stride_x_1*ctr_1 + _data_x_00;
for (int ctr_2 = 0; ctr_2 < _size_x_2; ctr_2 += 1)
{
const float d = b + c + e + _data_x_00_10[_stride_x_2*ctr_2];
_data_x_00_10[_stride_x_2*ctr_2] = b + c + _data_x_00_10[_stride_x_2*ctr_2];
}
}
}
}
```https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/54Always use codegen.rewriting.optimize2019-09-23T13:38:39+02:00Stephan SeitzAlways use codegen.rewriting.optimizePretty much !34 but with the changes to `create_kernel`. Can be closed if not wanted. Leaving it here for archiving purposes.
!34 has the workflow:
```python
assignments = optimize(assignments, optimizations)
ast = create_kernel(...Pretty much !34 but with the changes to `create_kernel`. Can be closed if not wanted. Leaving it here for archiving purposes.
!34 has the workflow:
```python
assignments = optimize(assignments, optimizations)
ast = create_kernel(assignments)
```https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/53Compile CUDA using the LLVM backend2019-09-23T12:49:30+02:00Stephan SeitzCompile CUDA using the LLVM backendWe can compile CUDA to PTX using the LLVM backend :wink:
`llc` produces PTX files without complaining.We can compile CUDA to PTX using the LLVM backend :wink:
`llc` produces PTX files without complaining.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/52Sort headers/global definitions to enable reproducible code generation2019-09-23T11:03:53+02:00Stephan SeitzSort headers/global definitions to enable reproducible code generationheaders and global_declarations are generated by methods that return
sets. So even with the same inputs it is not guaranteed that the same
source code is generated since sets do not guarantee a specific order
when iterating over them.
I...headers and global_declarations are generated by methods that return
sets. So even with the same inputs it is not guaranteed that the same
source code is generated since sets do not guarantee a specific order
when iterating over them.
I was supprised that my generated code could often not be reused from the cache. The problem was that the included headers appeared in random order.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/34Address #13: Use sympy.codegen.rewriting.optimize2019-09-23T10:55:13+02:00Stephan SeitzAddress #13: Use sympy.codegen.rewriting.optimizeIt's really comfortable to write optimizations in terms of `sympy.codegen.rewrite.RewriteOptim`:
```python
# Evaluates all constant terms
evaluate_constant_terms = ReplaceOptim(
lambda e: hasattr(e, 'is_constant') a...It's really comfortable to write optimizations in terms of `sympy.codegen.rewrite.RewriteOptim`:
```python
# Evaluates all constant terms
evaluate_constant_terms = ReplaceOptim(
lambda e: hasattr(e, 'is_constant') and e.is_constant,
lambda p: p.evalf()
)
```
This PR adds a parameter `sympy_optimizations` to the `create_*_kernel` functions that applies the list of optimizations to the assignments before creating the AST.
`sympy.codegen.rewrite` already has some optimizations. Some similar to the optimizations of pystencils.
For example `create_expand_pow_optimization(limit)` is really similar to the logic in `CustomSympyPrinter._print_Pow`.
See #13
Problem: old versions of sympy (e.g. from ubuntu CI) don't have `sympy.codegen.rewrite`. The optimizations are skipped in that case. `test_and_coverage` applies all optimizations.
We could also try to implement a fma-optimization (fused-multipy add) with that and `sympy.Wild`.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/46AES-NI vectorization improvements2019-09-17T09:08:05+02:00Michael Kuronmkuron@icp.uni-stuttgart.deAES-NI vectorization improvements!30 didn't implement an SSE-vectorized `_mm_cvtepu64_pd` equivalent because the [stackoverflow](https://stackoverflow.com/a/41148578) solution didn't work. That turned out to be due to a bad optimization in GCC 5+ in fast-math mode. None...!30 didn't implement an SSE-vectorized `_mm_cvtepu64_pd` equivalent because the [stackoverflow](https://stackoverflow.com/a/41148578) solution didn't work. That turned out to be due to a bad optimization in GCC 5+ in fast-math mode. None of the other compilers (Clang, Intel, MSVC) have that issue, so we just disable fast-math for that function.
Also, we now use fused multiply-add if available.Martin BauerMartin Bauerhttps://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/50CI: Add minimal-sympy-master2019-09-17T09:06:54+02:00Stephan SeitzCI: Add minimal-sympy-masterTest for #11
This could warn us when SymPy introduces breaking changes.
The test with SymPy master is allowed to fail.Test for #11
This could warn us when SymPy introduces breaking changes.
The test with SymPy master is allowed to fail.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/51Close pystencils' config file after writing2019-09-17T09:06:15+02:00Stephan SeitzClose pystencils' config file after writingI got a warning that this file remains unclosed.I got a warning that this file remains unclosed.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/49Actually increment counter inside random_symbol2019-09-04T15:06:57+02:00Michael Kuronmkuron@icp.uni-stuttgart.deActually increment counter inside random_symbol@rudolfweeber is currently looking at the statistical mechanics of the fluctuating LB and found a velocity bias. It turned out that this is due to all generated random numbers using the same key. Instead it should be incremented when gen...@rudolfweeber is currently looking at the statistical mechanics of the fluctuating LB and found a velocity bias. It turned out that this is due to all generated random numbers using the same key. Instead it should be incremented when generating multiple random numbers in the same kernel.
So in the generated code,
```c++
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 0, 2, Dummy_38, Dummy_39, Dummy_40, Dummy_41);
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 0, 2, Dummy_34, Dummy_35, Dummy_36, Dummy_37);
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 0, 2, Dummy_30, Dummy_31, Dummy_32, Dummy_33);
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 0, 2, Dummy_26, Dummy_27, Dummy_28, Dummy_29);
```
becomes
```c++
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 3, 2, Dummy_38, Dummy_39, Dummy_40, Dummy_41);
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 2, 2, Dummy_34, Dummy_35, Dummy_36, Dummy_37);
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 1, 2, Dummy_30, Dummy_31, Dummy_32, Dummy_33);
philox_float4(time_step, ctr_0, ctr_1, ctr_2, 0, 2, Dummy_26, Dummy_27, Dummy_28, Dummy_29);
```Martin BauerMartin Bauerhttps://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/41Add pyhtml to tests and artifacts2019-09-02T13:42:18+02:00Stephan SeitzAdd pyhtml to tests and artifactsI think this should suffice to produce the artifacts.
But `pytest-html` and `ansi2html` need to be added to the docker images.I think this should suffice to produce the artifacts.
But `pytest-html` and `ansi2html` need to be added to the docker images.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/44Fix typo in "pre-push"2019-09-02T13:41:09+02:00Stephan SeitzFix typo in "pre-push"https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/45Add PyPI badge2019-09-02T13:40:44+02:00Stephan SeitzAdd PyPI badgeBadge with current PyPI version and link to the PyPI page.Badge with current PyPI version and link to the PyPI page.https://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/30AES-NI Random Number Generator2019-09-02T10:21:21+02:00Michael Kuronmkuron@icp.uni-stuttgart.deAES-NI Random Number GeneratorI was looking at how to vectorize the Philox RNG yesterday. Before I knew it, I had implemented a working RNG using AES-NI instructions :nerd: ... Not entirely what I had intended to do, but it might still be useful to someone and should...I was looking at how to vectorize the Philox RNG yesterday. Before I knew it, I had implemented a working RNG using AES-NI instructions :nerd: ... Not entirely what I had intended to do, but it might still be useful to someone and should be similarly fast as a vectorized Philox.
There is one place that could be optimized because I fall back to scalar instructions: I failed to reimplement `_mm_cvtepu64_pd` (the solution from https://stackoverflow.com/a/41148578 produces incorrect results in the least-significant half of the mantissa). Perhaps someone else can try to fix that.
I did not integrate this with the `vector_instruction_set` parameter of the code generation. Perhaps you can do that, @bauer. It needs support for SSE2 and AES instructions (which look like SSE2 instructions, but their availability is determined by a separate CPUID flag). It will also make use of `_mm_cvtepu32_ps` and `_mm_cvtepu64_pd` from AVX512 if available (these are 128-bit instructions that actually look like SSE2 instructions).Martin BauerMartin Bauerhttps://i10git.cs.fau.de/pycodegen/pystencils/-/merge_requests/42WIP: make kerncraft/matplotlib tests pass with new image2019-09-02T08:28:51+02:00Stephan SeitzWIP: make kerncraft/matplotlib tests pass with new image--