ZTWHHH commited on
Commit
79fde12
·
verified ·
1 Parent(s): 55f6825

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. vglm/bin/python +3 -0
  3. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/_foreach_log2_native.h +25 -0
  4. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/_foreach_sqrt_native.h +25 -0
  5. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/grid_sampler_native.h +21 -0
  6. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/hstack_ops.h +39 -0
  7. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/native_layer_norm_cpu_dispatch.h +24 -0
  8. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/renorm_meta.h +27 -0
  9. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/sinh_cuda_dispatch.h +26 -0
  10. videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/sparse_resize_and_clear_ops.h +50 -0
  11. vllm/lib/python3.10/site-packages/cupyx/__pycache__/__init__.cpython-310.pyc +0 -0
  12. vllm/lib/python3.10/site-packages/cupyx/__pycache__/_rsqrt.cpython-310.pyc +0 -0
  13. vllm/lib/python3.10/site-packages/cupyx/__pycache__/_runtime.cpython-310.pyc +0 -0
  14. vllm/lib/python3.10/site-packages/cupyx/__pycache__/_scatter.cpython-310.pyc +0 -0
  15. vllm/lib/python3.10/site-packages/cupyx/__pycache__/time.cpython-310.pyc +0 -0
  16. vllm/lib/python3.10/site-packages/cupyx/scipy/__init__.py +35 -0
  17. vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/__init__.py +0 -0
  18. vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/__pycache__/__init__.cpython-310.pyc +0 -0
  19. vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/__pycache__/_util.cpython-310.pyc +0 -0
  20. vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/_util.py +73 -0
  21. vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/__init__.py +9 -0
  22. vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/__pycache__/__init__.cpython-310.pyc +0 -0
  23. vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/__pycache__/_fft.cpython-310.pyc +0 -0
  24. vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/_fft.py +508 -0
  25. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__init__.py +175 -0
  26. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_czt.cpython-310.pyc +0 -0
  27. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_filter_design.cpython-310.pyc +0 -0
  28. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_iir_filter_conversions.cpython-310.pyc +0 -0
  29. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_iir_filter_design.cpython-310.pyc +0 -0
  30. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_iir_utils.cpython-310.pyc +0 -0
  31. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_lti_conversion.cpython-310.pyc +0 -0
  32. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_ltisys.cpython-310.pyc +0 -0
  33. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_max_len_seq.cpython-310.pyc +0 -0
  34. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_resample.cpython-310.pyc +0 -0
  35. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_signaltools.cpython-310.pyc +0 -0
  36. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_signaltools_core.cpython-310.pyc +0 -0
  37. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_spectral_impl.cpython-310.pyc +0 -0
  38. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_splines.cpython-310.pyc +0 -0
  39. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_upfirdn.cpython-310.pyc +0 -0
  40. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_arraytools.py +272 -0
  41. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_bsplines.py +596 -0
  42. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_czt.py +451 -0
  43. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_filter_design.py +811 -0
  44. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_fir_filter_design.py +918 -0
  45. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_iir_filter_conversions.py +2364 -0
  46. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_iir_filter_design.py +994 -0
  47. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_iir_utils.py +737 -0
  48. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_lti_conversion.py +82 -0
  49. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_ltisys.py +0 -0
  50. vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_max_len_seq.py +119 -0
.gitattributes CHANGED
@@ -1438,3 +1438,4 @@ parrot/lib/python3.10/site-packages/numpy/_core/tests/__pycache__/test_numeric.c
1438
  parrot/lib/python3.10/site-packages/numpy/lib/tests/__pycache__/test_function_base.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
1439
  vllm/lib/python3.10/site-packages/huggingface_hub/inference/_generated/__pycache__/_async_client.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
1440
  vllm/lib/python3.10/site-packages/huggingface_hub/inference/__pycache__/_client.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
 
 
1438
  parrot/lib/python3.10/site-packages/numpy/lib/tests/__pycache__/test_function_base.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
1439
  vllm/lib/python3.10/site-packages/huggingface_hub/inference/_generated/__pycache__/_async_client.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
1440
  vllm/lib/python3.10/site-packages/huggingface_hub/inference/__pycache__/_client.cpython-310.pyc filter=lfs diff=lfs merge=lfs -text
1441
+ vglm/bin/python filter=lfs diff=lfs merge=lfs -text
vglm/bin/python ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5893c715ba63b749ab2bb330a28cd17a9e1f8ab65c7b580e714da9cec0caee3
3
+ size 17225608
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/_foreach_log2_native.h ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+
3
+ // @generated by torchgen/gen.py from NativeFunction.h
4
+
5
+ #include <c10/core/Scalar.h>
6
+ #include <c10/core/Storage.h>
7
+ #include <c10/core/TensorOptions.h>
8
+ #include <c10/util/Deprecated.h>
9
+ #include <c10/util/Optional.h>
10
+ #include <c10/core/QScheme.h>
11
+ #include <ATen/core/Reduction.h>
12
+ #include <ATen/core/Tensor.h>
13
+ #include <tuple>
14
+ #include <vector>
15
+
16
+
17
+ namespace at {
18
+ namespace native {
19
+ TORCH_API void _foreach_log2_out(at::TensorList self, at::TensorList out);
20
+ TORCH_API ::std::vector<at::Tensor> foreach_tensor_log2_slow(at::TensorList self);
21
+ TORCH_API void foreach_tensor_log2_slow_(at::TensorList self);
22
+ TORCH_API ::std::vector<at::Tensor> foreach_tensor_log2_cuda(at::TensorList self);
23
+ TORCH_API void foreach_tensor_log2_cuda_(at::TensorList self);
24
+ } // namespace native
25
+ } // namespace at
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/_foreach_sqrt_native.h ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+
3
+ // @generated by torchgen/gen.py from NativeFunction.h
4
+
5
+ #include <c10/core/Scalar.h>
6
+ #include <c10/core/Storage.h>
7
+ #include <c10/core/TensorOptions.h>
8
+ #include <c10/util/Deprecated.h>
9
+ #include <c10/util/Optional.h>
10
+ #include <c10/core/QScheme.h>
11
+ #include <ATen/core/Reduction.h>
12
+ #include <ATen/core/Tensor.h>
13
+ #include <tuple>
14
+ #include <vector>
15
+
16
+
17
+ namespace at {
18
+ namespace native {
19
+ TORCH_API void _foreach_sqrt_out(at::TensorList self, at::TensorList out);
20
+ TORCH_API ::std::vector<at::Tensor> foreach_tensor_sqrt_slow(at::TensorList self);
21
+ TORCH_API void foreach_tensor_sqrt_slow_(at::TensorList self);
22
+ TORCH_API ::std::vector<at::Tensor> foreach_tensor_sqrt_cuda(at::TensorList self);
23
+ TORCH_API void foreach_tensor_sqrt_cuda_(at::TensorList self);
24
+ } // namespace native
25
+ } // namespace at
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/grid_sampler_native.h ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+
3
+ // @generated by torchgen/gen.py from NativeFunction.h
4
+
5
+ #include <c10/core/Scalar.h>
6
+ #include <c10/core/Storage.h>
7
+ #include <c10/core/TensorOptions.h>
8
+ #include <c10/util/Deprecated.h>
9
+ #include <c10/util/Optional.h>
10
+ #include <c10/core/QScheme.h>
11
+ #include <ATen/core/Reduction.h>
12
+ #include <ATen/core/Tensor.h>
13
+ #include <tuple>
14
+ #include <vector>
15
+
16
+
17
+ namespace at {
18
+ namespace native {
19
+ TORCH_API at::Tensor grid_sampler(const at::Tensor & input, const at::Tensor & grid, int64_t interpolation_mode, int64_t padding_mode, bool align_corners);
20
+ } // namespace native
21
+ } // namespace at
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/hstack_ops.h ADDED
@@ -0,0 +1,39 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+
3
+ // @generated by torchgen/gen.py from Operator.h
4
+
5
+ #include <tuple>
6
+ #include <vector>
7
+
8
+ // Forward declarations of any types needed in the operator signatures.
9
+ // We can't directly include these classes because it will cause circular include dependencies.
10
+ // This file is included by TensorBody.h, which defines the Tensor class.
11
+ #include <ATen/core/ATen_fwd.h>
12
+
13
+ namespace at {
14
+ namespace _ops {
15
+
16
+
17
+ struct TORCH_API hstack {
18
+ using schema = at::Tensor (at::TensorList);
19
+ using ptr_schema = schema*;
20
+ // See Note [static constexpr char* members for windows NVCC]
21
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::hstack")
22
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "")
23
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "hstack(Tensor[] tensors) -> Tensor")
24
+ static at::Tensor call(at::TensorList tensors);
25
+ static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, at::TensorList tensors);
26
+ };
27
+
28
+ struct TORCH_API hstack_out {
29
+ using schema = at::Tensor & (at::TensorList, at::Tensor &);
30
+ using ptr_schema = schema*;
31
+ // See Note [static constexpr char* members for windows NVCC]
32
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::hstack")
33
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "out")
34
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "hstack.out(Tensor[] tensors, *, Tensor(a!) out) -> Tensor(a!)")
35
+ static at::Tensor & call(at::TensorList tensors, at::Tensor & out);
36
+ static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::TensorList tensors, at::Tensor & out);
37
+ };
38
+
39
+ }} // namespace at::_ops
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/native_layer_norm_cpu_dispatch.h ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+ // @generated by torchgen/gen.py from DispatchKeyFunction.h
3
+
4
+ // NB: The implementing C++ file is RegisterDispatchKey.cpp
5
+
6
+ // The only #includes we need are for custom classes that have defaults in the C++ API
7
+ #include <c10/core/MemoryFormat.h>
8
+ #include <c10/core/Scalar.h>
9
+ #include <ATen/core/Reduction.h>
10
+
11
+ // Forward declarations of any types needed in the operator signatures.
12
+ // We can't directly include these classes because it will cause circular include dependencies.
13
+ // This file is included by TensorBody.h, which defines the Tensor class.
14
+ #include <ATen/core/ATen_fwd.h>
15
+
16
+ namespace at {
17
+
18
+ namespace cpu {
19
+
20
+ TORCH_API ::std::tuple<at::Tensor,at::Tensor,at::Tensor> native_layer_norm(const at::Tensor & input, at::IntArrayRef normalized_shape, const c10::optional<at::Tensor> & weight, const c10::optional<at::Tensor> & bias, double eps);
21
+ TORCH_API ::std::tuple<at::Tensor,at::Tensor,at::Tensor> native_layer_norm_symint(const at::Tensor & input, c10::SymIntArrayRef normalized_shape, const c10::optional<at::Tensor> & weight, const c10::optional<at::Tensor> & bias, double eps);
22
+
23
+ } // namespace cpu
24
+ } // namespace at
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/renorm_meta.h ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+
3
+ // @generated by torchgen/gen.py from NativeMetaFunction.h
4
+
5
+ #include <c10/core/Scalar.h>
6
+ #include <c10/core/Storage.h>
7
+ #include <c10/core/TensorOptions.h>
8
+ #include <c10/util/Deprecated.h>
9
+ #include <c10/util/Optional.h>
10
+ #include <c10/core/QScheme.h>
11
+ #include <ATen/core/Reduction.h>
12
+ #include <ATen/TensorIterator.h>
13
+ #include <ATen/TensorMeta.h>
14
+ #include <tuple>
15
+ #include <vector>
16
+
17
+ namespace at {
18
+ namespace meta {
19
+
20
+ struct TORCH_API structured_renorm : public at::impl::MetaBase {
21
+
22
+
23
+ void meta(const at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm);
24
+ };
25
+
26
+ } // namespace native
27
+ } // namespace at
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/sinh_cuda_dispatch.h ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+ // @generated by torchgen/gen.py from DispatchKeyFunction.h
3
+
4
+ // NB: The implementing C++ file is RegisterDispatchKey.cpp
5
+
6
+ // The only #includes we need are for custom classes that have defaults in the C++ API
7
+ #include <c10/core/MemoryFormat.h>
8
+ #include <c10/core/Scalar.h>
9
+ #include <ATen/core/Reduction.h>
10
+
11
+ // Forward declarations of any types needed in the operator signatures.
12
+ // We can't directly include these classes because it will cause circular include dependencies.
13
+ // This file is included by TensorBody.h, which defines the Tensor class.
14
+ #include <ATen/core/ATen_fwd.h>
15
+
16
+ namespace at {
17
+
18
+ namespace cuda {
19
+
20
+ TORCH_API at::Tensor sinh(const at::Tensor & self);
21
+ TORCH_API at::Tensor & sinh_out(at::Tensor & out, const at::Tensor & self);
22
+ TORCH_API at::Tensor & sinh_outf(const at::Tensor & self, at::Tensor & out);
23
+ TORCH_API at::Tensor & sinh_(at::Tensor & self);
24
+
25
+ } // namespace cuda
26
+ } // namespace at
videollama2/lib/python3.10/site-packages/torch/include/ATen/ops/sparse_resize_and_clear_ops.h ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #pragma once
2
+
3
+ // @generated by torchgen/gen.py from Operator.h
4
+
5
+ #include <tuple>
6
+ #include <vector>
7
+
8
+ // Forward declarations of any types needed in the operator signatures.
9
+ // We can't directly include these classes because it will cause circular include dependencies.
10
+ // This file is included by TensorBody.h, which defines the Tensor class.
11
+ #include <ATen/core/ATen_fwd.h>
12
+
13
+ namespace at {
14
+ namespace _ops {
15
+
16
+
17
+ struct TORCH_API sparse_resize_and_clear_ {
18
+ using schema = const at::Tensor & (const at::Tensor &, at::IntArrayRef, int64_t, int64_t);
19
+ using ptr_schema = schema*;
20
+ // See Note [static constexpr char* members for windows NVCC]
21
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::sparse_resize_and_clear_")
22
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "")
23
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "sparse_resize_and_clear_(Tensor(a!) self, int[] size, int sparse_dim, int dense_dim) -> Tensor(a!)")
24
+ static const at::Tensor & call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim);
25
+ static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim);
26
+ };
27
+
28
+ struct TORCH_API sparse_resize_and_clear_out {
29
+ using schema = const at::Tensor & (const at::Tensor &, at::IntArrayRef, int64_t, int64_t, const at::Tensor &);
30
+ using ptr_schema = schema*;
31
+ // See Note [static constexpr char* members for windows NVCC]
32
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::sparse_resize_and_clear")
33
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "out")
34
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "sparse_resize_and_clear.out(Tensor self, int[] size, int sparse_dim, int dense_dim, *, Tensor(a!) out) -> Tensor(a!)")
35
+ static const at::Tensor & call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim, const at::Tensor & out);
36
+ static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim, const at::Tensor & out);
37
+ };
38
+
39
+ struct TORCH_API sparse_resize_and_clear {
40
+ using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, int64_t, int64_t);
41
+ using ptr_schema = schema*;
42
+ // See Note [static constexpr char* members for windows NVCC]
43
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(name, "aten::sparse_resize_and_clear")
44
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(overload_name, "")
45
+ STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(schema_str, "sparse_resize_and_clear(Tensor self, int[] size, int sparse_dim, int dense_dim) -> Tensor")
46
+ static at::Tensor call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim);
47
+ static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim);
48
+ };
49
+
50
+ }} // namespace at::_ops
vllm/lib/python3.10/site-packages/cupyx/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (1.19 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/__pycache__/_rsqrt.cpython-310.pyc ADDED
Binary file (347 Bytes). View file
 
vllm/lib/python3.10/site-packages/cupyx/__pycache__/_runtime.cpython-310.pyc ADDED
Binary file (8.4 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/__pycache__/_scatter.cpython-310.pyc ADDED
Binary file (5.27 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/__pycache__/time.cpython-310.pyc ADDED
Binary file (3.39 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/__init__.py ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import sys as _sys
2
+
3
+ from cupy._core import ndarray as _ndarray
4
+ from cupyx.scipy.sparse._base import spmatrix as _spmatrix
5
+
6
+
7
+ try:
8
+ import scipy as _scipy
9
+ _scipy_available = True
10
+ except ImportError:
11
+ _scipy_available = False
12
+
13
+
14
+ _cupyx_scipy = _sys.modules[__name__]
15
+
16
+
17
+ def get_array_module(*args):
18
+ """Returns the array module for arguments.
19
+
20
+ This function is used to implement CPU/GPU generic code. If at least one of
21
+ the arguments is a :class:`cupy.ndarray` object, the :mod:`cupyx.scipy`
22
+ module is returned.
23
+
24
+ Args:
25
+ args: Values to determine whether NumPy or CuPy should be used.
26
+
27
+ Returns:
28
+ module: :mod:`cupyx.scipy` or :mod:`scipy` is returned based on the
29
+ types of the arguments.
30
+
31
+ """
32
+ for arg in args:
33
+ if isinstance(arg, (_ndarray, _spmatrix)):
34
+ return _cupyx_scipy
35
+ return _scipy
vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/__init__.py ADDED
File without changes
vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (166 Bytes). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/__pycache__/_util.cpython-310.pyc ADDED
Binary file (2.17 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/_lib/_util.py ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+
3
+ import cupy
4
+
5
+
6
+ def float_factorial(n):
7
+ """Compute the factorial and return as a float
8
+
9
+ Returns infinity when result is too large for a double
10
+ """
11
+ return float(math.factorial(n)) if n < 171 else cupy.inf
12
+
13
+
14
+ def _asarray_validated(a, check_finite=True,
15
+ sparse_ok=False, objects_ok=False, mask_ok=False,
16
+ as_inexact=False):
17
+ """Helper function for SciPy argument validation.
18
+
19
+ Many CuPy linear algebra functions do support arbitrary array-like
20
+ input arguments. Examples of commonly unsupported inputs include
21
+ matrices containing inf/nan, sparse matrix representations, and
22
+ matrices with complicated elements.
23
+
24
+ Parameters
25
+ ----------
26
+ a : array-like
27
+ The array-like input
28
+ check_finite : bool, optional
29
+ By default True. To check whether the input matrices contain
30
+ only finite numbers. Disabling may give a performance gain,
31
+ but may result in problems (crashes, non-termination) if the
32
+ inputs do contain infinites or NaNs
33
+ sparse_ok : bool, optional
34
+ By default False. True if cupy sparse matrices are allowed
35
+ objects_ok : bool, optional
36
+ By default False. True if arrays with dype('O') are allowed
37
+ mask_ok : bool, optional
38
+ By default False. True if masked arrays are allowed.
39
+ as_inexact : bool, optional
40
+ By default False. True to convert the input array to a
41
+ cupy.inexact dtype
42
+
43
+ Returns
44
+ -------
45
+ ret : cupy.ndarray
46
+ The converted validated array
47
+
48
+ """
49
+
50
+ if not sparse_ok:
51
+ import cupyx.scipy.sparse
52
+ if cupyx.scipy.sparse.issparse(a):
53
+ msg = ('Sparse matrices are not supported by this function. '
54
+ 'Perhaps one of the cupyx.scipy.sparse.linalg functions '
55
+ 'would work instead.')
56
+ raise ValueError(msg)
57
+
58
+ # TODO: remove these comments when CuPy supports masked arrays
59
+ # Ref Issue: https://github.com/cupy/cupy/issues/2225
60
+ # if not mask_ok:
61
+ # if cupy.ma.isMaskedArray(a):
62
+ # raise ValueError('masked arrays are not supported')
63
+
64
+ # TODO: remove these comments when CuPy supports 'object' dtype
65
+ # if not objects_ok:
66
+ # if a.dtype is cupy.dtype('O'):
67
+ # raise ValueError('object arrays are not supported')
68
+
69
+ if as_inexact:
70
+ if not cupy.issubdtype(a, cupy.inexact):
71
+ a = a.astype(dtype=cupy.float64)
72
+
73
+ return a
vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/__init__.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ from cupyx.scipy.fftpack._fft import fft # NOQA
2
+ from cupyx.scipy.fftpack._fft import fft2 # NOQA
3
+ from cupyx.scipy.fftpack._fft import fftn # NOQA
4
+ from cupyx.scipy.fftpack._fft import ifft # NOQA
5
+ from cupyx.scipy.fftpack._fft import ifft2 # NOQA
6
+ from cupyx.scipy.fftpack._fft import ifftn # NOQA
7
+ from cupyx.scipy.fftpack._fft import irfft # NOQA
8
+ from cupyx.scipy.fftpack._fft import rfft # NOQA
9
+ from cupyx.scipy.fftpack._fft import get_fft_plan # NOQA
vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (448 Bytes). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/__pycache__/_fft.cpython-310.pyc ADDED
Binary file (16.8 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/fftpack/_fft.py ADDED
@@ -0,0 +1,508 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from numpy import prod
2
+
3
+ import cupy
4
+ from cupy.fft import config
5
+ from cupy.fft._fft import (_convert_fft_type, _default_fft_func, _fft,
6
+ _get_cufft_plan_nd, _get_fftn_out_size,
7
+ _output_dtype)
8
+ from cupy.fft._cache import get_plan_cache
9
+
10
+
11
+ def get_fft_plan(a, shape=None, axes=None, value_type='C2C'):
12
+ """ Generate a CUDA FFT plan for transforming up to three axes.
13
+
14
+ Args:
15
+ a (cupy.ndarray): Array to be transform, assumed to be either C- or
16
+ F- contiguous.
17
+ shape (None or tuple of ints): Shape of the transformed axes of the
18
+ output. If ``shape`` is not given, the lengths of the input along
19
+ the axes specified by ``axes`` are used.
20
+ axes (None or int or tuple of int): The axes of the array to
21
+ transform. If `None`, it is assumed that all axes are transformed.
22
+
23
+ Currently, for performing N-D transform these must be a set of up
24
+ to three adjacent axes, and must include either the first or the
25
+ last axis of the array.
26
+ value_type (str): The FFT type to perform. Acceptable values are:
27
+
28
+ * 'C2C': complex-to-complex transform (default)
29
+ * 'R2C': real-to-complex transform
30
+ * 'C2R': complex-to-real transform
31
+
32
+ Returns:
33
+ a cuFFT plan for either 1D transform (``cupy.cuda.cufft.Plan1d``) or
34
+ N-D transform (``cupy.cuda.cufft.PlanNd``).
35
+
36
+ .. note::
37
+ The returned plan can not only be passed as one of the arguments of
38
+ the functions in ``cupyx.scipy.fftpack``, but also be used as a
39
+ context manager for both ``cupy.fft`` and ``cupyx.scipy.fftpack``
40
+ functions:
41
+
42
+ .. code-block:: python
43
+
44
+ x = cupy.random.random(16).reshape(4, 4).astype(complex)
45
+ plan = cupyx.scipy.fftpack.get_fft_plan(x)
46
+ with plan:
47
+ y = cupy.fft.fftn(x)
48
+ # alternatively:
49
+ y = cupyx.scipy.fftpack.fftn(x) # no explicit plan is given!
50
+ # alternatively:
51
+ y = cupyx.scipy.fftpack.fftn(x, plan=plan) # pass plan explicitly
52
+
53
+ In the first case, no cuFFT plan will be generated automatically,
54
+ even if ``cupy.fft.config.enable_nd_planning = True`` is set.
55
+
56
+ .. note::
57
+ If this function is called under the context of
58
+ :func:`~cupy.fft.config.set_cufft_callbacks`, the generated plan will
59
+ have callbacks enabled.
60
+
61
+ .. warning::
62
+ This API is a deviation from SciPy's, is currently experimental, and
63
+ may be changed in the future version.
64
+ """
65
+ from cupy.cuda import cufft
66
+
67
+ # check input array
68
+ if a.flags.c_contiguous:
69
+ order = 'C'
70
+ elif a.flags.f_contiguous:
71
+ order = 'F'
72
+ else:
73
+ raise ValueError('Input array a must be contiguous')
74
+
75
+ if isinstance(shape, int):
76
+ shape = (shape,)
77
+ if isinstance(axes, int):
78
+ axes = (axes,)
79
+ if (shape is not None) and (axes is not None) and len(shape) != len(axes):
80
+ raise ValueError('Shape and axes have different lengths.')
81
+
82
+ # check axes
83
+ # n=1: 1d (need axis1D); n>1: Nd
84
+ if axes is None:
85
+ n = a.ndim if shape is None else len(shape)
86
+ axes = tuple(i for i in range(-n, 0))
87
+ if n == 1:
88
+ axis1D = 0
89
+ else: # axes is a tuple
90
+ n = len(axes)
91
+ if n == 1:
92
+ axis1D = axes[0]
93
+ if axis1D >= a.ndim or axis1D < -a.ndim:
94
+ err = 'The chosen axis ({0}) exceeds the number of '\
95
+ 'dimensions of a ({1})'.format(axis1D, a.ndim)
96
+ raise ValueError(err)
97
+ elif n > 3:
98
+ raise ValueError('Only up to three axes is supported')
99
+
100
+ # Note that "shape" here refers to the shape along transformed axes, not
101
+ # the shape of the output array, and we need to convert it to the latter.
102
+ # The result is as if "a=_cook_shape(a); return a.shape" is called.
103
+ # Because of this, we need to use (possibly unsorted) axes.
104
+ transformed_shape = shape
105
+ shape = list(a.shape)
106
+ if transformed_shape is not None:
107
+ for s, axis in zip(transformed_shape, axes):
108
+ if s is not None:
109
+ if axis == axes[-1] and value_type == 'C2R':
110
+ s = s // 2 + 1
111
+ shape[axis] = s
112
+ shape = tuple(shape)
113
+
114
+ # check value_type
115
+ out_dtype = _output_dtype(a.dtype, value_type)
116
+ fft_type = _convert_fft_type(out_dtype, value_type)
117
+ # TODO(leofang): figure out if we really have to skip F-order?
118
+ if n > 1 and value_type != 'C2C' and a.flags.f_contiguous:
119
+ raise ValueError('C2R/R2C PlanNd for F-order arrays is not supported')
120
+
121
+ # generate plan
122
+ # (load from cache if it exists, otherwise create one but don't cache it)
123
+ if n > 1: # ND transform
124
+ if cupy.cuda.runtime.is_hip and value_type == 'C2R':
125
+ raise RuntimeError("hipFFT's C2R PlanNd is buggy and unsupported")
126
+ out_size = _get_fftn_out_size(
127
+ shape, transformed_shape, axes[-1], value_type)
128
+ # _get_cufft_plan_nd interacts with plan cache and callback
129
+ plan = _get_cufft_plan_nd(
130
+ shape, fft_type, axes=axes, order=order, out_size=out_size,
131
+ to_cache=False)
132
+ else: # 1D transform
133
+ # prepare plan arguments
134
+ if value_type != 'C2R':
135
+ out_size = shape[axis1D]
136
+ else:
137
+ out_size = _get_fftn_out_size(
138
+ shape, transformed_shape, axis1D, value_type)
139
+ batch = prod(shape) // shape[axis1D]
140
+ devices = None if not config.use_multi_gpus else config._devices
141
+
142
+ keys = (out_size, fft_type, batch, devices)
143
+ mgr = config.get_current_callback_manager()
144
+ if mgr is not None:
145
+ # to avoid a weird segfault, we generate and cache distinct plans
146
+ # for every possible (load_aux, store_aux) pairs; the plans are
147
+ # still generated from the same external Python module
148
+ load_aux = mgr.cb_load_aux_arr
149
+ store_aux = mgr.cb_store_aux_arr
150
+ keys += (mgr.cb_load, mgr.cb_store,
151
+ 0 if load_aux is None else load_aux.data.ptr,
152
+ 0 if store_aux is None else store_aux.data.ptr)
153
+ cache = get_plan_cache()
154
+ cached_plan = cache.get(keys)
155
+ if cached_plan is not None:
156
+ plan = cached_plan
157
+ elif mgr is None:
158
+ plan = cufft.Plan1d(out_size, fft_type, batch, devices=devices)
159
+ else: # has callback
160
+ # TODO(leofang): support multi-GPU callback (devices is ignored)
161
+ if devices:
162
+ raise NotImplementedError('multi-GPU cuFFT callbacks are not '
163
+ 'yet supported')
164
+ plan = mgr.create_plan(('Plan1d', keys[:-3]))
165
+ mgr.set_callbacks(plan)
166
+
167
+ return plan
168
+
169
+
170
+ def fft(x, n=None, axis=-1, overwrite_x=False, plan=None):
171
+ """Compute the one-dimensional FFT.
172
+
173
+ Args:
174
+ x (cupy.ndarray): Array to be transformed.
175
+ n (None or int): Length of the transformed axis of the output. If ``n``
176
+ is not given, the length of the input along the axis specified by
177
+ ``axis`` is used.
178
+ axis (int): Axis over which to compute the FFT.
179
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
180
+ plan (:class:`cupy.cuda.cufft.Plan1d` or ``None``): a cuFFT plan for
181
+ transforming ``x`` over ``axis``, which can be obtained using::
182
+
183
+ plan = cupyx.scipy.fftpack.get_fft_plan(x, axis)
184
+
185
+ Note that `plan` is defaulted to None, meaning CuPy will use an
186
+ auto-generated plan behind the scene.
187
+
188
+ Returns:
189
+ cupy.ndarray:
190
+ The transformed array which shape is specified by ``n`` and type
191
+ will convert to complex if that of the input is another.
192
+
193
+ .. note::
194
+ The argument `plan` is currently experimental and the interface may be
195
+ changed in the future version.
196
+
197
+ .. seealso:: :func:`scipy.fftpack.fft`
198
+ """
199
+ from cupy.cuda import cufft
200
+ return _fft(x, (n,), (axis,), None, cufft.CUFFT_FORWARD,
201
+ overwrite_x=overwrite_x, plan=plan)
202
+
203
+
204
+ def ifft(x, n=None, axis=-1, overwrite_x=False, plan=None):
205
+ """Compute the one-dimensional inverse FFT.
206
+
207
+ Args:
208
+ x (cupy.ndarray): Array to be transformed.
209
+ n (None or int): Length of the transformed axis of the output. If ``n``
210
+ is not given, the length of the input along the axis specified by
211
+ ``axis`` is used.
212
+ axis (int): Axis over which to compute the FFT.
213
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
214
+ plan (:class:`cupy.cuda.cufft.Plan1d` or ``None``): a cuFFT plan for
215
+ transforming ``x`` over ``axis``, which can be obtained using::
216
+
217
+ plan = cupyx.scipy.fftpack.get_fft_plan(x, axis)
218
+
219
+ Note that `plan` is defaulted to None, meaning CuPy will use an
220
+ auto-generated plan behind the scene.
221
+
222
+ Returns:
223
+ cupy.ndarray:
224
+ The transformed array which shape is specified by ``n`` and type
225
+ will convert to complex if that of the input is another.
226
+
227
+ .. note::
228
+ The argument `plan` is currently experimental and the interface may be
229
+ changed in the future version.
230
+
231
+ .. seealso:: :func:`scipy.fftpack.ifft`
232
+ """
233
+ from cupy.cuda import cufft
234
+ return _fft(x, (n,), (axis,), None, cufft.CUFFT_INVERSE,
235
+ overwrite_x=overwrite_x, plan=plan)
236
+
237
+
238
+ def fft2(x, shape=None, axes=(-2, -1), overwrite_x=False, plan=None):
239
+ """Compute the two-dimensional FFT.
240
+
241
+ Args:
242
+ x (cupy.ndarray): Array to be transformed.
243
+ shape (None or tuple of ints): Shape of the transformed axes of the
244
+ output. If ``shape`` is not given, the lengths of the input along
245
+ the axes specified by ``axes`` are used.
246
+ axes (tuple of ints): Axes over which to compute the FFT.
247
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
248
+ plan (:class:`cupy.cuda.cufft.PlanNd` or ``None``): a cuFFT plan for
249
+ transforming ``x`` over ``axes``, which can be obtained using::
250
+
251
+ plan = cupyx.scipy.fftpack.get_fft_plan(x, axes)
252
+
253
+ Note that `plan` is defaulted to None, meaning CuPy will either
254
+ use an auto-generated plan behind the scene if cupy.fft.config.
255
+ enable_nd_planning = True, or use no cuFFT plan if it is set to
256
+ False.
257
+
258
+ Returns:
259
+ cupy.ndarray:
260
+ The transformed array which shape is specified by ``shape`` and
261
+ type will convert to complex if that of the input is another.
262
+
263
+ .. seealso:: :func:`scipy.fftpack.fft2`
264
+
265
+ .. note::
266
+ The argument `plan` is currently experimental and the interface may be
267
+ changed in the future version.
268
+ """
269
+ from cupy.cuda import cufft
270
+
271
+ func = _default_fft_func(x, shape, axes, plan)
272
+ return func(x, shape, axes, None, cufft.CUFFT_FORWARD,
273
+ overwrite_x=overwrite_x, plan=plan)
274
+
275
+
276
+ def ifft2(x, shape=None, axes=(-2, -1), overwrite_x=False, plan=None):
277
+ """Compute the two-dimensional inverse FFT.
278
+
279
+ Args:
280
+ x (cupy.ndarray): Array to be transformed.
281
+ shape (None or tuple of ints): Shape of the transformed axes of the
282
+ output. If ``shape`` is not given, the lengths of the input along
283
+ the axes specified by ``axes`` are used.
284
+ axes (tuple of ints): Axes over which to compute the FFT.
285
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
286
+ plan (:class:`cupy.cuda.cufft.PlanNd` or ``None``): a cuFFT plan for
287
+ transforming ``x`` over ``axes``, which can be obtained using::
288
+
289
+ plan = cupyx.scipy.fftpack.get_fft_plan(x, axes)
290
+
291
+ Note that `plan` is defaulted to None, meaning CuPy will either
292
+ use an auto-generated plan behind the scene if cupy.fft.config.
293
+ enable_nd_planning = True, or use no cuFFT plan if it is set to
294
+ False.
295
+
296
+ Returns:
297
+ cupy.ndarray:
298
+ The transformed array which shape is specified by ``shape`` and
299
+ type will convert to complex if that of the input is another.
300
+
301
+ .. seealso:: :func:`scipy.fftpack.ifft2`
302
+
303
+ .. note::
304
+ The argument `plan` is currently experimental and the interface may be
305
+ changed in the future version.
306
+ """
307
+ from cupy.cuda import cufft
308
+
309
+ func = _default_fft_func(x, shape, axes, plan)
310
+ return func(x, shape, axes, None, cufft.CUFFT_INVERSE,
311
+ overwrite_x=overwrite_x, plan=plan)
312
+
313
+
314
+ def fftn(x, shape=None, axes=None, overwrite_x=False, plan=None):
315
+ """Compute the N-dimensional FFT.
316
+
317
+ Args:
318
+ x (cupy.ndarray): Array to be transformed.
319
+ shape (None or tuple of ints): Shape of the transformed axes of the
320
+ output. If ``shape`` is not given, the lengths of the input along
321
+ the axes specified by ``axes`` are used.
322
+ axes (tuple of ints): Axes over which to compute the FFT.
323
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
324
+ plan (:class:`cupy.cuda.cufft.PlanNd` or ``None``): a cuFFT plan for
325
+ transforming ``x`` over ``axes``, which can be obtained using::
326
+
327
+ plan = cupyx.scipy.fftpack.get_fft_plan(x, axes)
328
+
329
+ Note that `plan` is defaulted to None, meaning CuPy will either
330
+ use an auto-generated plan behind the scene if cupy.fft.config.
331
+ enable_nd_planning = True, or use no cuFFT plan if it is set to
332
+ False.
333
+
334
+ Returns:
335
+ cupy.ndarray:
336
+ The transformed array which shape is specified by ``shape`` and
337
+ type will convert to complex if that of the input is another.
338
+
339
+ .. seealso:: :func:`scipy.fftpack.fftn`
340
+
341
+ .. note::
342
+ The argument `plan` is currently experimental and the interface may be
343
+ changed in the future version.
344
+ """
345
+ from cupy.cuda import cufft
346
+
347
+ func = _default_fft_func(x, shape, axes, plan)
348
+ return func(x, shape, axes, None, cufft.CUFFT_FORWARD,
349
+ overwrite_x=overwrite_x, plan=plan)
350
+
351
+
352
+ def ifftn(x, shape=None, axes=None, overwrite_x=False, plan=None):
353
+ """Compute the N-dimensional inverse FFT.
354
+
355
+ Args:
356
+ x (cupy.ndarray): Array to be transformed.
357
+ shape (None or tuple of ints): Shape of the transformed axes of the
358
+ output. If ``shape`` is not given, the lengths of the input along
359
+ the axes specified by ``axes`` are used.
360
+ axes (tuple of ints): Axes over which to compute the FFT.
361
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
362
+ plan (:class:`cupy.cuda.cufft.PlanNd` or ``None``): a cuFFT plan for
363
+ transforming ``x`` over ``axes``, which can be obtained using::
364
+
365
+ plan = cupyx.scipy.fftpack.get_fft_plan(x, axes)
366
+
367
+ Note that `plan` is defaulted to None, meaning CuPy will either
368
+ use an auto-generated plan behind the scene if cupy.fft.config.
369
+ enable_nd_planning = True, or use no cuFFT plan if it is set to
370
+ False.
371
+
372
+ Returns:
373
+ cupy.ndarray:
374
+ The transformed array which shape is specified by ``shape`` and
375
+ type will convert to complex if that of the input is another.
376
+
377
+ .. seealso:: :func:`scipy.fftpack.ifftn`
378
+
379
+ .. note::
380
+ The argument `plan` is currently experimental and the interface may be
381
+ changed in the future version.
382
+ """
383
+ from cupy.cuda import cufft
384
+
385
+ func = _default_fft_func(x, shape, axes, plan)
386
+ return func(x, shape, axes, None, cufft.CUFFT_INVERSE,
387
+ overwrite_x=overwrite_x, plan=plan)
388
+
389
+
390
+ def rfft(x, n=None, axis=-1, overwrite_x=False, plan=None):
391
+ """Compute the one-dimensional FFT for real input.
392
+
393
+ The returned real array contains
394
+
395
+ .. code-block:: python
396
+
397
+ [y(0),Re(y(1)),Im(y(1)),...,Re(y(n/2))] # if n is even
398
+ [y(0),Re(y(1)),Im(y(1)),...,Re(y(n/2)),Im(y(n/2))] # if n is odd
399
+
400
+ Args:
401
+ x (cupy.ndarray): Array to be transformed.
402
+ n (None or int): Length of the transformed axis of the output. If ``n``
403
+ is not given, the length of the input along the axis specified by
404
+ ``axis`` is used.
405
+ axis (int): Axis over which to compute the FFT.
406
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
407
+ plan (:class:`cupy.cuda.cufft.Plan1d` or ``None``): a cuFFT plan for
408
+ transforming ``x`` over ``axis``, which can be obtained using::
409
+
410
+ plan = cupyx.scipy.fftpack.get_fft_plan(
411
+ x, axes, value_type='R2C')
412
+
413
+ Note that `plan` is defaulted to None, meaning CuPy will either
414
+ use an auto-generated plan behind the scene if cupy.fft.config.
415
+ enable_nd_planning = True, or use no cuFFT plan if it is set to
416
+ False.
417
+
418
+ Returns:
419
+ cupy.ndarray:
420
+ The transformed array.
421
+
422
+ .. seealso:: :func:`scipy.fftpack.rfft`
423
+
424
+ .. note::
425
+ The argument `plan` is currently experimental and the interface may be
426
+ changed in the future version.
427
+ """
428
+ from cupy.cuda import cufft
429
+
430
+ if n is None:
431
+ n = x.shape[axis]
432
+
433
+ shape = list(x.shape)
434
+ shape[axis] = n
435
+ f = _fft(x, (n,), (axis,), None, cufft.CUFFT_FORWARD, 'R2C',
436
+ overwrite_x=overwrite_x, plan=plan)
437
+ z = cupy.empty(shape, f.real.dtype)
438
+
439
+ slice_z = [slice(None)] * x.ndim
440
+ slice_f = [slice(None)] * x.ndim
441
+
442
+ slice_z[axis] = slice(1)
443
+ slice_f[axis] = slice(1)
444
+ z[tuple(slice_z)] = f[tuple(slice_f)].real
445
+
446
+ slice_z[axis] = slice(1, None, 2)
447
+ slice_f[axis] = slice(1, None)
448
+ z[tuple(slice_z)] = f[tuple(slice_f)].real
449
+
450
+ slice_z[axis] = slice(2, None, 2)
451
+ slice_f[axis] = slice(1, n - f.shape[axis] + 1)
452
+ z[tuple(slice_z)] = f[tuple(slice_f)].imag
453
+
454
+ return z
455
+
456
+
457
+ def irfft(x, n=None, axis=-1, overwrite_x=False):
458
+ """Compute the one-dimensional inverse FFT for real input.
459
+
460
+ Args:
461
+ x (cupy.ndarray): Array to be transformed.
462
+ n (None or int): Length of the transformed axis of the output. If ``n``
463
+ is not given, the length of the input along the axis specified by
464
+ ``axis`` is used.
465
+ axis (int): Axis over which to compute the FFT.
466
+ overwrite_x (bool): If True, the contents of ``x`` can be destroyed.
467
+
468
+ Returns:
469
+ cupy.ndarray:
470
+ The transformed array.
471
+
472
+ .. seealso:: :func:`scipy.fftpack.irfft`
473
+
474
+ .. note::
475
+ This function does not support a precomputed `plan`. If you need this
476
+ capability, please consider using :func:`cupy.fft.irfft` or :func:`
477
+ cupyx.scipy.fft.irfft`.
478
+ """
479
+ from cupy.cuda import cufft
480
+
481
+ if n is None:
482
+ n = x.shape[axis]
483
+ m = min(n, x.shape[axis])
484
+
485
+ shape = list(x.shape)
486
+ shape[axis] = n // 2 + 1
487
+ if x.dtype in (cupy.float16, cupy.float32):
488
+ z = cupy.zeros(shape, dtype=cupy.complex64)
489
+ else:
490
+ z = cupy.zeros(shape, dtype=cupy.complex128)
491
+
492
+ slice_x = [slice(None)] * x.ndim
493
+ slice_z = [slice(None)] * x.ndim
494
+
495
+ slice_x[axis] = slice(1)
496
+ slice_z[axis] = slice(1)
497
+ z[tuple(slice_z)].real = x[tuple(slice_x)]
498
+
499
+ slice_x[axis] = slice(1, m, 2)
500
+ slice_z[axis] = slice(1, m // 2 + 1)
501
+ z[tuple(slice_z)].real = x[tuple(slice_x)]
502
+
503
+ slice_x[axis] = slice(2, m, 2)
504
+ slice_z[axis] = slice(1, (m + 1) // 2)
505
+ z[tuple(slice_z)].imag = x[tuple(slice_x)]
506
+
507
+ return _fft(z, (n,), (axis,), None, cufft.CUFFT_INVERSE, 'C2R',
508
+ overwrite_x=overwrite_x)
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__init__.py ADDED
@@ -0,0 +1,175 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from cupyx.scipy.signal._signaltools import convolve # NOQA
2
+ from cupyx.scipy.signal._signaltools import correlate # NOQA
3
+ from cupyx.scipy.signal._signaltools import deconvolve # NOQA
4
+ from cupyx.scipy.signal._signaltools import fftconvolve # NOQA
5
+ from cupyx.scipy.signal._signaltools import choose_conv_method # NOQA
6
+ from cupyx.scipy.signal._signaltools import oaconvolve # NOQA
7
+ from cupyx.scipy.signal._signaltools import convolve2d # NOQA
8
+ from cupyx.scipy.signal._signaltools import correlate2d # NOQA
9
+ from cupyx.scipy.signal._signaltools import correlation_lags # NOQA
10
+ from cupyx.scipy.signal._signaltools import wiener # NOQA
11
+ from cupyx.scipy.signal._signaltools import order_filter # NOQA
12
+ from cupyx.scipy.signal._signaltools import medfilt # NOQA
13
+ from cupyx.scipy.signal._signaltools import medfilt2d # NOQA
14
+ from cupyx.scipy.signal._signaltools import lfilter # NOQA
15
+ from cupyx.scipy.signal._signaltools import lfiltic # NOQA
16
+ from cupyx.scipy.signal._signaltools import lfilter_zi # NOQA
17
+ from cupyx.scipy.signal._signaltools import detrend # NOQA
18
+ from cupyx.scipy.signal._signaltools import filtfilt # NOQA
19
+ from cupyx.scipy.signal._signaltools import sosfilt # NOQA
20
+ from cupyx.scipy.signal._signaltools import sosfilt_zi # NOQA
21
+ from cupyx.scipy.signal._signaltools import sosfiltfilt # NOQA
22
+ from cupyx.scipy.signal._signaltools import hilbert # NOQA
23
+ from cupyx.scipy.signal._signaltools import hilbert2 # NOQA
24
+
25
+ from cupyx.scipy.signal._resample import resample # NOQA
26
+ from cupyx.scipy.signal._resample import resample_poly # NOQA
27
+ from cupyx.scipy.signal._resample import decimate # NOQA
28
+
29
+ from cupyx.scipy.signal._polyutils import unique_roots # NOQA
30
+ from cupyx.scipy.signal._polyutils import invres # NOQA
31
+ from cupyx.scipy.signal._polyutils import invresz # NOQA
32
+ from cupyx.scipy.signal._polyutils import residue # NOQA
33
+ from cupyx.scipy.signal._polyutils import residuez # NOQA
34
+
35
+ from cupyx.scipy.signal._bsplines import sepfir2d # NOQA
36
+ from cupyx.scipy.signal._bsplines import cspline1d # NOQA
37
+ from cupyx.scipy.signal._bsplines import qspline1d # NOQA
38
+ from cupyx.scipy.signal._bsplines import cspline2d # NOQA
39
+ from cupyx.scipy.signal._bsplines import qspline2d # NOQA
40
+ from cupyx.scipy.signal._bsplines import cspline1d_eval # NOQA
41
+ from cupyx.scipy.signal._bsplines import qspline1d_eval # NOQA
42
+ from cupyx.scipy.signal._bsplines import spline_filter # NOQA
43
+ from cupyx.scipy.signal._bsplines import gauss_spline # NOQA
44
+
45
+ from cupyx.scipy.signal._splines import symiirorder1 # NOQA
46
+ from cupyx.scipy.signal._splines import symiirorder2 # NOQA
47
+
48
+ from cupyx.scipy.signal._savitzky_golay import savgol_coeffs, savgol_filter # NOQA
49
+
50
+ from cupyx.scipy.signal._filter_design import gammatone # NOQA
51
+ from cupyx.scipy.signal._filter_design import group_delay # NOQA
52
+
53
+ from cupyx.scipy.signal._fir_filter_design import kaiser_atten # NOQA
54
+ from cupyx.scipy.signal._fir_filter_design import kaiser_beta # NOQA
55
+ from cupyx.scipy.signal._fir_filter_design import kaiserord # NOQA
56
+
57
+ from cupyx.scipy.signal._iir_filter_conversions import BadCoefficients # NOQA
58
+ from cupyx.scipy.signal._iir_filter_conversions import normalize # NOQA
59
+
60
+ from cupyx.scipy.signal._iir_filter_conversions import bilinear # NOQA
61
+ from cupyx.scipy.signal._iir_filter_conversions import lp2lp # NOQA
62
+ from cupyx.scipy.signal._iir_filter_conversions import lp2hp # NOQA
63
+ from cupyx.scipy.signal._iir_filter_conversions import lp2bp # NOQA
64
+ from cupyx.scipy.signal._iir_filter_conversions import lp2bs # NOQA
65
+
66
+ from cupyx.scipy.signal._iir_filter_conversions import bilinear_zpk # NOQA
67
+ from cupyx.scipy.signal._iir_filter_conversions import lp2lp_zpk # NOQA
68
+ from cupyx.scipy.signal._iir_filter_conversions import lp2hp_zpk # NOQA
69
+ from cupyx.scipy.signal._iir_filter_conversions import lp2bp_zpk # NOQA
70
+ from cupyx.scipy.signal._iir_filter_conversions import lp2bs_zpk # NOQA
71
+
72
+ from cupyx.scipy.signal._iir_filter_conversions import zpk2tf # NOQA
73
+ from cupyx.scipy.signal._iir_filter_conversions import zpk2sos # NOQA
74
+ from cupyx.scipy.signal._iir_filter_conversions import zpk2ss # NOQA
75
+ from cupyx.scipy.signal._iir_filter_conversions import tf2zpk # NOQA
76
+ from cupyx.scipy.signal._iir_filter_conversions import tf2sos # NOQA
77
+ from cupyx.scipy.signal._iir_filter_conversions import tf2ss # NOQA
78
+ from cupyx.scipy.signal._iir_filter_conversions import ss2tf # NOQA
79
+ from cupyx.scipy.signal._iir_filter_conversions import ss2zpk # NOQA
80
+ from cupyx.scipy.signal._iir_filter_conversions import sos2tf # NOQA
81
+ from cupyx.scipy.signal._iir_filter_conversions import sos2zpk # NOQA
82
+
83
+ from cupyx.scipy.signal._iir_filter_conversions import band_stop_obj # NOQA
84
+ from cupyx.scipy.signal.windows._windows import get_window # NOQA
85
+
86
+ from cupyx.scipy.signal._iir_filter_conversions import buttap # NOQA
87
+ from cupyx.scipy.signal._iir_filter_conversions import cheb1ap # NOQA
88
+ from cupyx.scipy.signal._iir_filter_conversions import cheb2ap # NOQA
89
+ from cupyx.scipy.signal._iir_filter_conversions import ellipap # NOQA
90
+
91
+ from cupyx.scipy.signal._iir_filter_conversions import buttord # NOQA
92
+ from cupyx.scipy.signal._iir_filter_conversions import cheb1ord # NOQA
93
+ from cupyx.scipy.signal._iir_filter_conversions import cheb2ord # NOQA
94
+ from cupyx.scipy.signal._iir_filter_conversions import ellipord # NOQA
95
+
96
+ from cupyx.scipy.signal._iir_filter_design import iirfilter # NOQA
97
+ from cupyx.scipy.signal._iir_filter_design import butter # NOQA
98
+ from cupyx.scipy.signal._iir_filter_design import cheby1 # NOQA
99
+ from cupyx.scipy.signal._iir_filter_design import cheby2 # NOQA
100
+ from cupyx.scipy.signal._iir_filter_design import ellip # NOQA
101
+ from cupyx.scipy.signal._iir_filter_design import iirdesign # NOQA
102
+ from cupyx.scipy.signal._iir_filter_design import iircomb # NOQA
103
+ from cupyx.scipy.signal._iir_filter_design import iirnotch # NOQA
104
+ from cupyx.scipy.signal._iir_filter_design import iirpeak # NOQA
105
+
106
+ from cupyx.scipy.signal._fir_filter_design import firwin # NOQA
107
+ from cupyx.scipy.signal._fir_filter_design import firwin2 # NOQA
108
+ from cupyx.scipy.signal._fir_filter_design import firls # NOQA
109
+ from cupyx.scipy.signal._fir_filter_design import minimum_phase # NOQA
110
+
111
+ from cupyx.scipy.signal._filter_design import findfreqs # NOQA
112
+ from cupyx.scipy.signal._filter_design import freqs # NOQA
113
+ from cupyx.scipy.signal._filter_design import freqs_zpk # NOQA
114
+
115
+ from cupyx.scipy.signal._filter_design import freqz # NOQA
116
+ from cupyx.scipy.signal._filter_design import freqz_zpk # NOQA
117
+ from cupyx.scipy.signal._filter_design import sosfreqz # NOQA
118
+
119
+ from cupyx.scipy.signal._waveforms import chirp # NOQA
120
+ from cupyx.scipy.signal._waveforms import gausspulse # NOQA
121
+ from cupyx.scipy.signal._waveforms import sawtooth # NOQA
122
+ from cupyx.scipy.signal._waveforms import square # NOQA
123
+ from cupyx.scipy.signal._waveforms import unit_impulse # NOQA
124
+ from cupyx.scipy.signal._max_len_seq import max_len_seq # NOQA
125
+
126
+ from cupyx.scipy.signal._czt import * # NOQA
127
+
128
+ from cupyx.scipy.signal._wavelets import morlet # NOQA
129
+ from cupyx.scipy.signal._wavelets import qmf # NOQA
130
+ from cupyx.scipy.signal._wavelets import ricker # NOQA
131
+ from cupyx.scipy.signal._wavelets import morlet2 # NOQA
132
+ from cupyx.scipy.signal._wavelets import cwt # NOQA
133
+
134
+ from cupyx.scipy.signal._lti_conversion import abcd_normalize # NOQA
135
+
136
+ from cupyx.scipy.signal._upfirdn import upfirdn # NOQA
137
+
138
+ from cupyx.scipy.signal._peak_finding import find_peaks # NOQA
139
+ from cupyx.scipy.signal._peak_finding import peak_prominences # NOQA
140
+ from cupyx.scipy.signal._peak_finding import peak_widths # NOQA
141
+
142
+ from cupyx.scipy.signal._ltisys import lti # NOQA
143
+ from cupyx.scipy.signal._ltisys import lsim # NOQA
144
+ from cupyx.scipy.signal._ltisys import impulse # NOQA
145
+ from cupyx.scipy.signal._ltisys import step # NOQA
146
+ from cupyx.scipy.signal._ltisys import freqresp # NOQA
147
+ from cupyx.scipy.signal._ltisys import bode # NOQA
148
+
149
+ from cupyx.scipy.signal._ltisys import dlti # NOQA
150
+ from cupyx.scipy.signal._ltisys import dlsim # NOQA
151
+ from cupyx.scipy.signal._ltisys import dstep # NOQA
152
+ from cupyx.scipy.signal._ltisys import dimpulse # NOQA
153
+ from cupyx.scipy.signal._ltisys import dbode # NOQA
154
+ from cupyx.scipy.signal._ltisys import dfreqresp # NOQA
155
+ from cupyx.scipy.signal._ltisys import StateSpace # NOQA
156
+ from cupyx.scipy.signal._ltisys import TransferFunction # NOQA
157
+ from cupyx.scipy.signal._ltisys import ZerosPolesGain # NOQA
158
+ from cupyx.scipy.signal._ltisys import cont2discrete # NOQA
159
+ from cupyx.scipy.signal._ltisys import place_poles # NOQA
160
+
161
+ from cupyx.scipy.signal._spectral import lombscargle # NOQA
162
+ from cupyx.scipy.signal._spectral import periodogram # NOQA
163
+ from cupyx.scipy.signal._spectral import welch # NOQA
164
+ from cupyx.scipy.signal._spectral import csd # NOQA
165
+ from cupyx.scipy.signal._spectral import check_COLA # NOQA
166
+ from cupyx.scipy.signal._spectral import check_NOLA # NOQA
167
+ from cupyx.scipy.signal._spectral import stft # NOQA
168
+ from cupyx.scipy.signal._spectral import istft # NOQA
169
+ from cupyx.scipy.signal._spectral import spectrogram # NOQA
170
+ from cupyx.scipy.signal._spectral import vectorstrength # NOQA
171
+ from cupyx.scipy.signal._spectral import coherence # NOQA
172
+
173
+ from cupyx.scipy.signal._peak_finding import argrelextrema # NOQA
174
+ from cupyx.scipy.signal._peak_finding import argrelmin # NOQA
175
+ from cupyx.scipy.signal._peak_finding import argrelmax # NOQA
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_czt.cpython-310.pyc ADDED
Binary file (15.3 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_filter_design.cpython-310.pyc ADDED
Binary file (23.9 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_iir_filter_conversions.cpython-310.pyc ADDED
Binary file (60.3 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_iir_filter_design.cpython-310.pyc ADDED
Binary file (31.6 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_iir_utils.cpython-310.pyc ADDED
Binary file (19.4 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_lti_conversion.cpython-310.pyc ADDED
Binary file (2.21 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_ltisys.cpython-310.pyc ADDED
Binary file (83.7 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_max_len_seq.cpython-310.pyc ADDED
Binary file (4.15 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_resample.cpython-310.pyc ADDED
Binary file (16.5 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_signaltools.cpython-310.pyc ADDED
Binary file (54.3 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_signaltools_core.cpython-310.pyc ADDED
Binary file (8.66 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_spectral_impl.cpython-310.pyc ADDED
Binary file (19 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_splines.cpython-310.pyc ADDED
Binary file (12.1 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/__pycache__/_upfirdn.cpython-310.pyc ADDED
Binary file (19.7 kB). View file
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_arraytools.py ADDED
@@ -0,0 +1,272 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Functions for acting on a axis of an array.
3
+ """
4
+ import cupy
5
+
6
+
7
+ def axis_slice(a, start=None, stop=None, step=None, axis=-1):
8
+ """Take a slice along axis 'axis' from 'a'.
9
+
10
+ Parameters
11
+ ----------
12
+ a : cupy.ndarray
13
+ The array to be sliced.
14
+ start, stop, step : int or None
15
+ The slice parameters.
16
+ axis : int, optional
17
+ The axis of `a` to be sliced.
18
+
19
+ Examples
20
+ --------
21
+ >>> a = array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
22
+ >>> axis_slice(a, start=0, stop=1, axis=1)
23
+ array([[1],
24
+ [4],
25
+ [7]])
26
+ >>> axis_slice(a, start=1, axis=0)
27
+ array([[4, 5, 6],
28
+ [7, 8, 9]])
29
+
30
+ Notes
31
+ -----
32
+ The keyword arguments start, stop and step are used by calling
33
+ slice(start, stop, step). This implies axis_slice() does not
34
+ handle its arguments the exactly the same as indexing. To select
35
+ a single index k, for example, use
36
+ axis_slice(a, start=k, stop=k+1)
37
+ In this case, the length of the axis 'axis' in the result will
38
+ be 1; the trivial dimension is not removed. (Use cupy.squeeze()
39
+ to remove trivial axes.)
40
+ """
41
+ a_slice = [slice(None)] * a.ndim
42
+ a_slice[axis] = slice(start, stop, step)
43
+ b = a[tuple(a_slice)]
44
+ return b
45
+
46
+
47
+ def axis_assign(a, b, start=None, stop=None, step=None, axis=-1):
48
+ """Take a slice along axis 'axis' from 'a' and set it to 'b' in-place.
49
+
50
+ Parameters
51
+ ----------
52
+ a : numpy.ndarray
53
+ The array to be sliced.
54
+ b : cupy.ndarray
55
+ The array to be assigned.
56
+ start, stop, step : int or None
57
+ The slice parameters.
58
+ axis : int, optional
59
+ The axis of `a` to be sliced.
60
+
61
+ Examples
62
+ --------
63
+ >>> a = array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
64
+ >>> b1 = array([[-1], [-4], [-7]])
65
+ >>> axis_assign(a, b1, start=0, stop=1, axis=1)
66
+ array([[-1, 2, 3],
67
+ [-4, 5, 6],
68
+ [-7, 8, 9]])
69
+
70
+ Notes
71
+ -----
72
+ The keyword arguments start, stop and step are used by calling
73
+ slice(start, stop, step). This implies axis_assign() does not
74
+ handle its arguments the exactly the same as indexing. To assign
75
+ a single index k, for example, use
76
+ axis_assign(a, start=k, stop=k+1)
77
+ In this case, the length of the axis 'axis' in the result will
78
+ be 1; the trivial dimension is not removed. (Use numpy.squeeze()
79
+ to remove trivial axes.)
80
+
81
+ This function works in-place and will modify the values contained in `a`
82
+ """
83
+ a_slice = [slice(None)] * a.ndim
84
+ a_slice[axis] = slice(start, stop, step)
85
+ a[tuple(a_slice)] = b
86
+ return a
87
+
88
+
89
+ def axis_reverse(a, axis=-1):
90
+ """Reverse the 1-D slices of `a` along axis `axis`.
91
+
92
+ Returns axis_slice(a, step=-1, axis=axis).
93
+ """
94
+ return axis_slice(a, step=-1, axis=axis)
95
+
96
+
97
+ def odd_ext(x, n, axis=-1):
98
+ """
99
+ Odd extension at the boundaries of an array
100
+
101
+ Generate a new ndarray by making an odd extension of `x` along an axis.
102
+
103
+ Parameters
104
+ ----------
105
+ x : ndarray
106
+ The array to be extended.
107
+ n : int
108
+ The number of elements by which to extend `x` at each end of the axis.
109
+ axis : int, optional
110
+ The axis along which to extend `x`. Default is -1.
111
+
112
+ Examples
113
+ --------
114
+ >>> from cupyx.scipy.signal._arraytools import odd_ext
115
+ >>> a = cupy.array([[1, 2, 3, 4, 5], [0, 1, 4, 9, 16]])
116
+ >>> odd_ext(a, 2)
117
+ array([[-1, 0, 1, 2, 3, 4, 5, 6, 7],
118
+ [-4, -1, 0, 1, 4, 9, 16, 23, 28]])
119
+ """
120
+ if n < 1:
121
+ return x
122
+ if n > x.shape[axis] - 1:
123
+ raise ValueError(("The extension length n (%d) is too big. " +
124
+ "It must not exceed x.shape[axis]-1, which is %d.")
125
+ % (n, x.shape[axis] - 1))
126
+ left_end = axis_slice(x, start=0, stop=1, axis=axis)
127
+ left_ext = axis_slice(x, start=n, stop=0, step=-1, axis=axis)
128
+ right_end = axis_slice(x, start=-1, axis=axis)
129
+ right_ext = axis_slice(x, start=-2, stop=-(n + 2), step=-1, axis=axis)
130
+ ext = cupy.concatenate((2 * left_end - left_ext, x,
131
+ 2 * right_end - right_ext), axis=axis)
132
+ return ext
133
+
134
+
135
+ def even_ext(x, n, axis=-1):
136
+ """
137
+ Even extension at the boundaries of an array
138
+
139
+ Generate a new ndarray by making an even extension of `x` along an axis.
140
+
141
+ Parameters
142
+ ----------
143
+ x : ndarray
144
+ The array to be extended.
145
+ n : int
146
+ The number of elements by which to extend `x` at each end of the axis.
147
+ axis : int, optional
148
+ The axis along which to extend `x`. Default is -1.
149
+
150
+ Examples
151
+ --------
152
+ >>> from cupyx.scipy.signal._arraytools import even_ext
153
+ >>> a = cupy.array([[1, 2, 3, 4, 5], [0, 1, 4, 9, 16]])
154
+ >>> even_ext(a, 2)
155
+ array([[ 3, 2, 1, 2, 3, 4, 5, 4, 3],
156
+ [ 4, 1, 0, 1, 4, 9, 16, 9, 4]])
157
+
158
+ """
159
+ if n < 1:
160
+ return x
161
+ if n > x.shape[axis] - 1:
162
+ raise ValueError(("The extension length n (%d) is too big. " +
163
+ "It must not exceed x.shape[axis]-1, which is %d.")
164
+ % (n, x.shape[axis] - 1))
165
+ left_ext = axis_slice(x, start=n, stop=0, step=-1, axis=axis)
166
+ right_ext = axis_slice(x, start=-2, stop=-(n + 2), step=-1, axis=axis)
167
+ ext = cupy.concatenate((left_ext, x, right_ext), axis=axis)
168
+ return ext
169
+
170
+
171
+ def const_ext(x, n, axis=-1):
172
+ """
173
+ Constant extension at the boundaries of an array
174
+
175
+ Generate a new ndarray that is a constant extension of `x` along an axis.
176
+ The extension repeats the values at the first and last element of
177
+ the axis.
178
+
179
+ Parameters
180
+ ----------
181
+ x : ndarray
182
+ The array to be extended.
183
+ n : int
184
+ The number of elements by which to extend `x` at each end of the axis.
185
+ axis : int, optional
186
+ The axis along which to extend `x`. Default is -1.
187
+
188
+ Examples
189
+ --------
190
+ >>> from cupyx.scipy.signal._arraytools import const_ext
191
+ >>> a = cupy.array([[1, 2, 3, 4, 5], [0, 1, 4, 9, 16]])
192
+ >>> const_ext(a, 2)
193
+ array([[ 1, 1, 1, 2, 3, 4, 5, 5, 5],
194
+ [ 0, 0, 0, 1, 4, 9, 16, 16, 16]])
195
+ """
196
+ if n < 1:
197
+ return x
198
+ left_end = axis_slice(x, start=0, stop=1, axis=axis)
199
+ ones_shape = [1] * x.ndim
200
+ ones_shape[axis] = n
201
+ ones = cupy.ones(ones_shape, dtype=x.dtype)
202
+ left_ext = ones * left_end
203
+ right_end = axis_slice(x, start=-1, axis=axis)
204
+ right_ext = ones * right_end
205
+ ext = cupy.concatenate((left_ext, x, right_ext), axis=axis)
206
+ return ext
207
+
208
+
209
+ def zero_ext(x, n, axis=-1):
210
+ """
211
+ Zero padding at the boundaries of an array
212
+
213
+ Generate a new ndarray that is a zero-padded extension of `x` along
214
+ an axis.
215
+
216
+ Parameters
217
+ ----------
218
+ x : ndarray
219
+ The array to be extended.
220
+ n : int
221
+ The number of elements by which to extend `x` at each end of the
222
+ axis.
223
+ axis : int, optional
224
+ The axis along which to extend `x`. Default is -1.
225
+
226
+ Examples
227
+ --------
228
+ >>> from cupyx.scipy.signal._arraytools import zero_ext
229
+ >>> a = cupy.array([[1, 2, 3, 4, 5], [0, 1, 4, 9, 16]])
230
+ >>> zero_ext(a, 2)
231
+ array([[ 0, 0, 1, 2, 3, 4, 5, 0, 0],
232
+ [ 0, 0, 0, 1, 4, 9, 16, 0, 0]])
233
+ """
234
+ if n < 1:
235
+ return x
236
+ zeros_shape = list(x.shape)
237
+ zeros_shape[axis] = n
238
+ zeros = cupy.zeros(zeros_shape, dtype=x.dtype)
239
+ ext = cupy.concatenate((zeros, x, zeros), axis=axis)
240
+ return ext
241
+
242
+
243
+ def _as_strided(x, shape=None, strides=None):
244
+ """
245
+ Create a view into the array with the given shape and strides.
246
+ .. warning:: This function has to be used with extreme care, see notes.
247
+
248
+ Parameters
249
+ ----------
250
+ x : ndarray
251
+ Array to create a new.
252
+ shape : sequence of int, optional
253
+ The shape of the new array. Defaults to ``x.shape``.
254
+ strides : sequence of int, optional
255
+ The strides of the new array. Defaults to ``x.strides``.
256
+
257
+ Returns
258
+ -------
259
+ view : ndarray
260
+
261
+ Notes
262
+ -----
263
+ ``as_strided`` creates a view into the array given the exact strides
264
+ and shape. This means it manipulates the internal data structure of
265
+ ndarray and, if done incorrectly, the array elements can point to
266
+ invalid memory and can corrupt results or crash your program.
267
+ """
268
+ shape = x.shape if shape is None else tuple(shape)
269
+ strides = x.strides if strides is None else tuple(strides)
270
+
271
+ return cupy.ndarray(
272
+ shape=shape, dtype=x.dtype, memptr=x.data, strides=strides)
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_bsplines.py ADDED
@@ -0,0 +1,596 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ """
3
+ Signal processing B-Splines
4
+
5
+ Some of the functions defined here were ported directly from CuSignal under
6
+ terms of the MIT license, under the following notice:
7
+
8
+ Copyright (c) 2019-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
9
+
10
+ Permission is hereby granted, free of charge, to any person obtaining a
11
+ copy of this software and associated documentation files (the "Software"),
12
+ to deal in the Software without restriction, including without limitation
13
+ the rights to use, copy, modify, merge, publish, distribute, sublicense,
14
+ and/or sell copies of the Software, and to permit persons to whom the
15
+ Software is furnished to do so, subject to the following conditions:
16
+
17
+ The above copyright notice and this permission notice shall be included in
18
+ all copies or substantial portions of the Software.
19
+
20
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
21
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
22
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
23
+ THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
24
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
25
+ FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
26
+ DEALINGS IN THE SOFTWARE.
27
+ """
28
+
29
+ import cupy
30
+ import cupyx.scipy.ndimage
31
+
32
+ from cupyx.scipy.signal._iir_utils import apply_iir_sos
33
+ from cupyx.scipy.signal._splines import _symiirorder1_nd, _symiirorder2_nd
34
+ from cupyx.scipy.interpolate._bspline import BSpline
35
+
36
+ import numpy as np
37
+
38
+
39
+ def sepfir2d(input, hrow, hcol):
40
+ """Convolve with a 2-D separable FIR filter.
41
+
42
+ Convolve the rank-2 input array with the separable filter defined by the
43
+ rank-1 arrays hrow, and hcol. Mirror symmetric boundary conditions are
44
+ assumed. This function can be used to find an image given its B-spline
45
+ representation.
46
+
47
+ The arguments `hrow` and `hcol` must be 1-dimensional and of off length.
48
+
49
+ Args:
50
+ input (cupy.ndarray): The input signal
51
+ hrow (cupy.ndarray): Row direction filter
52
+ hcol (cupy.ndarray): Column direction filter
53
+
54
+ Returns:
55
+ cupy.ndarray: The filtered signal
56
+
57
+ .. seealso:: :func:`scipy.signal.sepfir2d`
58
+ """
59
+ if any(x.ndim != 1 or x.size % 2 == 0 for x in (hrow, hcol)):
60
+ raise ValueError('hrow and hcol must be 1 dimensional and odd length')
61
+ dtype = input.dtype
62
+ if dtype.kind == 'c':
63
+ dtype = cupy.complex64 if dtype == cupy.complex64 else cupy.complex128
64
+ elif dtype == cupy.float32 or dtype.itemsize <= 2:
65
+ dtype = cupy.float32
66
+ else:
67
+ dtype = cupy.float64
68
+ input = input.astype(dtype, copy=False)
69
+ hrow = hrow.astype(dtype, copy=False)
70
+ hcol = hcol.astype(dtype, copy=False)
71
+ filters = (hcol[::-1].conj(), hrow[::-1].conj())
72
+ return cupyx.scipy.ndimage._filters._run_1d_correlates(
73
+ input, (0, 1), lambda i: filters[i], None, 'reflect', 0)
74
+
75
+
76
+ def _quadratic(x):
77
+ x = abs(cupy.asarray(x, dtype=float))
78
+ b = BSpline.basis_element(
79
+ cupy.asarray([-1.5, -0.5, 0.5, 1.5]), extrapolate=False)
80
+ out = b(x)
81
+ out[(x < -1.5) | (x > 1.5)] = 0
82
+ return out
83
+
84
+
85
+ def _cubic(x):
86
+ x = cupy.asarray(x, dtype=float)
87
+ b = BSpline.basis_element(
88
+ cupy.asarray([-2, -1, 0, 1, 2]), extrapolate=False)
89
+ out = b(x)
90
+ out[(x < -2) | (x > 2)] = 0
91
+ return out
92
+
93
+
94
+ @cupy.fuse()
95
+ def _coeff_smooth(lam):
96
+ xi = 1 - 96 * lam + 24 * lam * cupy.sqrt(3 + 144 * lam)
97
+ omeg = cupy.arctan2(cupy.sqrt(144 * lam - 1), cupy.sqrt(xi))
98
+ rho = (24 * lam - 1 - cupy.sqrt(xi)) / (24 * lam)
99
+ rho = rho * cupy.sqrt(
100
+ (48 * lam + 24 * lam * cupy.sqrt(3 + 144 * lam)) / xi)
101
+ return rho, omeg
102
+
103
+
104
+ @cupy.fuse()
105
+ def _hc(k, cs, rho, omega):
106
+ return (cs / cupy.sin(omega) * (rho ** k) * cupy.sin(omega * (k + 1)) *
107
+ cupy.greater(k, -1))
108
+
109
+
110
+ @cupy.fuse()
111
+ def _hs(k, cs, rho, omega):
112
+ c0 = (cs * cs * (1 + rho * rho) / (1 - rho * rho) /
113
+ (1 - 2 * rho * rho * cupy.cos(2 * omega) + rho ** 4))
114
+ gamma = (1 - rho * rho) / (1 + rho * rho) / cupy.tan(omega)
115
+ ak = cupy.abs(k)
116
+ return c0 * rho ** ak * (
117
+ cupy.cos(omega * ak) + gamma * cupy.sin(omega * ak))
118
+
119
+
120
+ def _cubic_smooth_coeff(signal, lamb):
121
+ rho, omega = _coeff_smooth(lamb)
122
+ cs = 1 - 2 * rho * cupy.cos(omega) + rho * rho
123
+ K = len(signal)
124
+ yp = cupy.zeros((K,), signal.dtype.char)
125
+ k = cupy.arange(K)
126
+
127
+ state_0 = (_hc(0, cs, rho, omega) * signal[0] +
128
+ cupy.sum(_hc(k + 1, cs, rho, omega) * signal))
129
+ state_1 = (_hc(0, cs, rho, omega) * signal[0] +
130
+ _hc(1, cs, rho, omega) * signal[1] +
131
+ cupy.sum(_hc(k + 2, cs, rho, omega) * signal))
132
+
133
+ zi = cupy.r_[0, 0, state_0, state_1]
134
+ zi = cupy.atleast_2d(zi)
135
+
136
+ coef = cupy.r_[cs, 0, 0, 1, -2 * rho * cupy.cos(omega), rho * rho]
137
+ coef = cupy.atleast_2d(coef)
138
+
139
+ # Forward pass:
140
+ #
141
+ # yp[0] = (_hc(0, cs, rho, omega) * signal[0] +
142
+ # cupy.sum(_hc(k + 1, cs, rho, omega) * signal))
143
+ # yp[1] = (_hc(0, cs, rho, omega) * signal[0] +
144
+ # _hc(1, cs, rho, omega) * signal[1] +
145
+ # cupy.sum(_hc(k + 2, cs, rho, omega) * signal))
146
+ # for n in range(2, K):
147
+ # yp[n] = (cs * signal[n] + 2 * rho * cupy.cos(omega) * yp[n - 1] -
148
+ # rho * rho * yp[n - 2])
149
+
150
+ yp, _ = apply_iir_sos(signal[2:], coef, zi=zi, dtype=signal.dtype)
151
+ yp = cupy.r_[state_0, state_1, yp]
152
+
153
+ # Reverse pass:
154
+ #
155
+ # y[K - 1] = cupy.sum((_hs(k, cs, rho, omega) +
156
+ # _hs(k + 1, cs, rho, omega)) * signal[::-1])
157
+ # y[K - 2] = cupy.sum((_hs(k - 1, cs, rho, omega) +
158
+ # _hs(k + 2, cs, rho, omega)) * signal[::-1])
159
+ # for n in range(K - 3, -1, -1):
160
+ # y[n] = (cs * yp[n] + 2 * rho * cupy.cos(omega) * y[n + 1] -
161
+ # rho * rho * y[n + 2])
162
+
163
+ state_0 = cupy.sum((_hs(k, cs, rho, omega) +
164
+ _hs(k + 1, cs, rho, omega)) * signal[::-1])
165
+ state_1 = cupy.sum((_hs(k - 1, cs, rho, omega) +
166
+ _hs(k + 2, cs, rho, omega)) * signal[::-1])
167
+
168
+ zi = cupy.r_[0, 0, state_0, state_1]
169
+ zi = cupy.atleast_2d(zi)
170
+
171
+ y, _ = apply_iir_sos(yp[-3::-1], coef, zi=zi, dtype=signal.dtype)
172
+ y = cupy.r_[y[::-1], state_1, state_0]
173
+ return y
174
+
175
+
176
+ def _cubic_coeff(signal):
177
+ zi = -2 + cupy.sqrt(3)
178
+ K = len(signal)
179
+ powers = zi ** cupy.arange(K)
180
+
181
+ if K == 1:
182
+ yplus = signal[0] + zi * cupy.sum(powers * signal)
183
+ output = zi / (zi - 1) * yplus
184
+ return cupy.atleast_1d(output)
185
+
186
+ state = cupy.r_[0, 0, 0, cupy.sum(powers * signal)]
187
+ state = cupy.atleast_2d(state)
188
+ coef = cupy.r_[1, 0, 0, 1, -zi, 0]
189
+ coef = cupy.atleast_2d(coef)
190
+
191
+ # yplus[0] = signal[0] + zi * sum(powers * signal)
192
+ # for k in range(1, K):
193
+ # yplus[k] = signal[k] + zi * yplus[k - 1]
194
+ yplus, _ = apply_iir_sos(signal, coef, zi=state, apply_fir=False,
195
+ dtype=signal.dtype)
196
+
197
+ out_last = zi / (zi - 1) * yplus[K - 1]
198
+ state = cupy.r_[0, 0, 0, out_last]
199
+ state = cupy.atleast_2d(state)
200
+
201
+ coef = cupy.r_[-zi, 0, 0, 1, -zi, 0]
202
+ coef = cupy.atleast_2d(coef)
203
+
204
+ # output[K - 1] = zi / (zi - 1) * yplus[K - 1]
205
+ # for k in range(K - 2, -1, -1):
206
+ # output[k] = zi * (output[k + 1] - yplus[k])
207
+ output, _ = apply_iir_sos(
208
+ yplus[-2::-1], coef, zi=state, dtype=signal.dtype)
209
+ output = cupy.r_[output[::-1], out_last]
210
+ return output * 6.0
211
+
212
+
213
+ def _quadratic_coeff(signal):
214
+ zi = -3 + 2 * cupy.sqrt(2.0)
215
+ K = len(signal)
216
+ powers = zi ** cupy.arange(K)
217
+
218
+ if K == 1:
219
+ yplus = signal[0] + zi * cupy.sum(powers * signal)
220
+ output = zi / (zi - 1) * yplus
221
+ return cupy.atleast_1d(output)
222
+
223
+ state = cupy.r_[0, 0, 0, cupy.sum(powers * signal)]
224
+ state = cupy.atleast_2d(state)
225
+ coef = cupy.r_[1, 0, 0, 1, -zi, 0]
226
+ coef = cupy.atleast_2d(coef)
227
+
228
+ # yplus[0] = signal[0] + zi * cupy.sum(powers * signal)
229
+ # for k in range(1, K):
230
+ # yplus[k] = signal[k] + zi * yplus[k - 1]
231
+ yplus, _ = apply_iir_sos(signal, coef, zi=state, apply_fir=False,
232
+ dtype=signal.dtype)
233
+
234
+ out_last = zi / (zi - 1) * yplus[K - 1]
235
+ state = cupy.r_[0, 0, 0, out_last]
236
+ state = cupy.atleast_2d(state)
237
+
238
+ coef = cupy.r_[-zi, 0, 0, 1, -zi, 0]
239
+ coef = cupy.atleast_2d(coef)
240
+
241
+ # output[K - 1] = zi / (zi - 1) * yplus[K - 1]
242
+ # for k in range(K - 2, -1, -1):
243
+ # output[k] = zi * (output[k + 1] - yplus[k])
244
+ output, _ = apply_iir_sos(
245
+ yplus[-2::-1], coef, zi=state, dtype=signal.dtype)
246
+ output = cupy.r_[output[::-1], out_last]
247
+ return output * 8.0
248
+
249
+
250
+ def compute_root_from_lambda(lamb):
251
+ tmp = np.sqrt(3 + 144 * lamb)
252
+ xi = 1 - 96 * lamb + 24 * lamb * tmp
253
+ omega = np.arctan(np.sqrt((144 * lamb - 1.0) / xi))
254
+ tmp2 = np.sqrt(xi)
255
+ r = ((24 * lamb - 1 - tmp2) / (24 * lamb) *
256
+ np.sqrt((48*lamb + 24 * lamb * tmp)) / tmp2)
257
+ return r, omega
258
+
259
+
260
+ def cspline1d(signal, lamb=0.0):
261
+ """
262
+ Compute cubic spline coefficients for rank-1 array.
263
+
264
+ Find the cubic spline coefficients for a 1-D signal assuming
265
+ mirror-symmetric boundary conditions. To obtain the signal back from the
266
+ spline representation mirror-symmetric-convolve these coefficients with a
267
+ length 3 FIR window [1.0, 4.0, 1.0]/ 6.0 .
268
+
269
+ Parameters
270
+ ----------
271
+ signal : ndarray
272
+ A rank-1 array representing samples of a signal.
273
+ lamb : float, optional
274
+ Smoothing coefficient, default is 0.0.
275
+
276
+ Returns
277
+ -------
278
+ c : ndarray
279
+ Cubic spline coefficients.
280
+
281
+ See Also
282
+ --------
283
+ cspline1d_eval : Evaluate a cubic spline at the new set of points.
284
+
285
+ """
286
+ if lamb != 0.0:
287
+ return _cubic_smooth_coeff(signal, lamb)
288
+ else:
289
+ return _cubic_coeff(signal)
290
+
291
+
292
+ def qspline1d(signal, lamb=0.0):
293
+ """Compute quadratic spline coefficients for rank-1 array.
294
+
295
+ Parameters
296
+ ----------
297
+ signal : ndarray
298
+ A rank-1 array representing samples of a signal.
299
+ lamb : float, optional
300
+ Smoothing coefficient (must be zero for now).
301
+
302
+ Returns
303
+ -------
304
+ c : ndarray
305
+ Quadratic spline coefficients.
306
+
307
+ See Also
308
+ --------
309
+ qspline1d_eval : Evaluate a quadratic spline at the new set of points.
310
+
311
+ Notes
312
+ -----
313
+ Find the quadratic spline coefficients for a 1-D signal assuming
314
+ mirror-symmetric boundary conditions. To obtain the signal back from the
315
+ spline representation mirror-symmetric-convolve these coefficients with a
316
+ length 3 FIR window [1.0, 6.0, 1.0]/ 8.0 .
317
+
318
+ """
319
+ if lamb != 0.0:
320
+ raise ValueError("Smoothing quadratic splines not supported yet.")
321
+ else:
322
+ return _quadratic_coeff(signal)
323
+
324
+
325
+ def cspline1d_eval(cj, newx, dx=1.0, x0=0):
326
+ """Evaluate a cubic spline at the new set of points.
327
+
328
+ `dx` is the old sample-spacing while `x0` was the old origin. In
329
+ other-words the old-sample points (knot-points) for which the `cj`
330
+ represent spline coefficients were at equally-spaced points of:
331
+
332
+ oldx = x0 + j*dx j=0...N-1, with N=len(cj)
333
+
334
+ Edges are handled using mirror-symmetric boundary conditions.
335
+
336
+ Parameters
337
+ ----------
338
+ cj : ndarray
339
+ cublic spline coefficients
340
+ newx : ndarray
341
+ New set of points.
342
+ dx : float, optional
343
+ Old sample-spacing, the default value is 1.0.
344
+ x0 : int, optional
345
+ Old origin, the default value is 0.
346
+
347
+ Returns
348
+ -------
349
+ res : ndarray
350
+ Evaluated a cubic spline points.
351
+
352
+ See Also
353
+ --------
354
+ cspline1d : Compute cubic spline coefficients for rank-1 array.
355
+
356
+ """
357
+ newx = (cupy.asarray(newx) - x0) / float(dx)
358
+ res = cupy.zeros_like(newx, dtype=cj.dtype)
359
+ if res.size == 0:
360
+ return res
361
+ N = len(cj)
362
+ cond1 = newx < 0
363
+ cond2 = newx > (N - 1)
364
+ cond3 = ~(cond1 | cond2)
365
+ # handle general mirror-symmetry
366
+ res[cond1] = cspline1d_eval(cj, -newx[cond1])
367
+ res[cond2] = cspline1d_eval(cj, 2 * (N - 1) - newx[cond2])
368
+ newx = newx[cond3]
369
+ if newx.size == 0:
370
+ return res
371
+ result = cupy.zeros_like(newx, dtype=cj.dtype)
372
+ jlower = cupy.floor(newx - 2).astype(int) + 1
373
+ for i in range(4):
374
+ thisj = jlower + i
375
+ indj = thisj.clip(0, N - 1) # handle edge cases
376
+ result += cj[indj] * _cubic(newx - thisj)
377
+ res[cond3] = result
378
+ return res
379
+
380
+
381
+ def qspline1d_eval(cj, newx, dx=1.0, x0=0):
382
+ """Evaluate a quadratic spline at the new set of points.
383
+
384
+ Parameters
385
+ ----------
386
+ cj : ndarray
387
+ Quadratic spline coefficients
388
+ newx : ndarray
389
+ New set of points.
390
+ dx : float, optional
391
+ Old sample-spacing, the default value is 1.0.
392
+ x0 : int, optional
393
+ Old origin, the default value is 0.
394
+
395
+ Returns
396
+ -------
397
+ res : ndarray
398
+ Evaluated a quadratic spline points.
399
+
400
+ See Also
401
+ --------
402
+ qspline1d : Compute quadratic spline coefficients for rank-1 array.
403
+
404
+ Notes
405
+ -----
406
+ `dx` is the old sample-spacing while `x0` was the old origin. In
407
+ other-words the old-sample points (knot-points) for which the `cj`
408
+ represent spline coefficients were at equally-spaced points of::
409
+
410
+ oldx = x0 + j*dx j=0...N-1, with N=len(cj)
411
+
412
+ Edges are handled using mirror-symmetric boundary conditions.
413
+
414
+ """
415
+ newx = (cupy.asarray(newx) - x0) / dx
416
+ res = cupy.zeros_like(newx)
417
+ if res.size == 0:
418
+ return res
419
+ N = len(cj)
420
+ cond1 = newx < 0
421
+ cond2 = newx > (N - 1)
422
+ cond3 = ~(cond1 | cond2)
423
+ # handle general mirror-symmetry
424
+ res[cond1] = qspline1d_eval(cj, -newx[cond1])
425
+ res[cond2] = qspline1d_eval(cj, 2 * (N - 1) - newx[cond2])
426
+ newx = newx[cond3]
427
+ if newx.size == 0:
428
+ return res
429
+ result = cupy.zeros_like(newx)
430
+ jlower = cupy.floor(newx - 1.5).astype(int) + 1
431
+ for i in range(3):
432
+ thisj = jlower + i
433
+ indj = thisj.clip(0, N - 1) # handle edge cases
434
+ result += cj[indj] * _quadratic(newx - thisj)
435
+ res[cond3] = result
436
+ return res
437
+
438
+
439
+ def cspline2d(signal, lamb=0.0, precision=-1.0):
440
+ """
441
+ Coefficients for 2-D cubic (3rd order) B-spline.
442
+
443
+ Return the third-order B-spline coefficients over a regularly spaced
444
+ input grid for the two-dimensional input image.
445
+
446
+ Parameters
447
+ ----------
448
+ input : ndarray
449
+ The input signal.
450
+ lamb : float
451
+ Specifies the amount of smoothing in the transfer function.
452
+ precision : float
453
+ Specifies the precision for computing the infinite sum needed to apply
454
+ mirror-symmetric boundary conditions.
455
+
456
+ Returns
457
+ -------
458
+ output : ndarray
459
+ The filtered signal.
460
+ """
461
+ if lamb <= 1 / 144.0:
462
+ # Normal cubic spline
463
+ r = -2 + np.sqrt(3.0)
464
+ out = _symiirorder1_nd(signal, -r * 6.0, r, precision=precision,
465
+ axis=-1)
466
+ out = _symiirorder1_nd(out, -r * 6.0, r, precision=precision,
467
+ axis=0)
468
+ return out
469
+
470
+ r, omega = compute_root_from_lambda(lamb)
471
+ out = _symiirorder2_nd(signal, r, omega, precision=precision, axis=-1)
472
+ out = _symiirorder2_nd(out, r, omega, precision=precision, axis=0)
473
+ return out
474
+
475
+
476
+ def qspline2d(signal, lamb=0.0, precision=-1.0):
477
+ """
478
+ Coefficients for 2-D quadratic (2nd order) B-spline.
479
+
480
+ Return the second-order B-spline coefficients over a regularly spaced
481
+ input grid for the two-dimensional input image.
482
+
483
+ Parameters
484
+ ----------
485
+ input : ndarray
486
+ The input signal.
487
+ lamb : float
488
+ Specifies the amount of smoothing in the transfer function.
489
+ precision : float
490
+ Specifies the precision for computing the infinite sum needed to apply
491
+ mirror-symmetric boundary conditions.
492
+
493
+ Returns
494
+ -------
495
+ output : ndarray
496
+ The filtered signal.
497
+ """
498
+
499
+ if lamb > 0:
500
+ raise ValueError('lambda must be negative or zero')
501
+
502
+ # normal quadratic spline
503
+ r = -3 + 2 * np.sqrt(2.0)
504
+
505
+ out = _symiirorder1_nd(signal, -r * 8.0, r, precision=precision, axis=-1)
506
+ out = _symiirorder1_nd(out, -r * 8.0, r, precision=precision, axis=0)
507
+ return out
508
+
509
+
510
+ def spline_filter(Iin, lmbda=5.0):
511
+ """Smoothing spline (cubic) filtering of a rank-2 array.
512
+
513
+ Filter an input data set, `Iin`, using a (cubic) smoothing spline of
514
+ fall-off `lmbda`.
515
+
516
+ Parameters
517
+ ----------
518
+ Iin : array_like
519
+ input data set
520
+ lmbda : float, optional
521
+ spline smooghing fall-off value, default is `5.0`.
522
+
523
+ Returns
524
+ -------
525
+ res : ndarray
526
+ filtered input data
527
+
528
+ """
529
+ intype = Iin.dtype.char
530
+ hcol = cupy.asarray([1.0, 4.0, 1.0], 'f') / 6.0
531
+ if intype in ['F', 'D']:
532
+ Iin = Iin.astype('F')
533
+ ckr = cspline2d(Iin.real, lmbda)
534
+ cki = cspline2d(Iin.imag, lmbda)
535
+ outr = sepfir2d(ckr, hcol, hcol)
536
+ outi = sepfir2d(cki, hcol, hcol)
537
+ out = (outr + 1j * outi).astype(intype)
538
+ elif intype in ['f', 'd']:
539
+ ckr = cspline2d(Iin, lmbda)
540
+ out = sepfir2d(ckr, hcol, hcol)
541
+ out = out.astype(intype)
542
+ else:
543
+ raise TypeError("Invalid data type for Iin")
544
+ return out
545
+
546
+
547
+ _gauss_spline_kernel = cupy.ElementwiseKernel(
548
+ "T x, int32 n",
549
+ "T output",
550
+ """
551
+ output = 1 / sqrt( 2.0 * M_PI * signsq ) * exp( -( x * x ) * r_signsq );
552
+ """,
553
+ "_gauss_spline_kernel",
554
+ options=("-std=c++11",),
555
+ loop_prep="const double signsq { ( n + 1 ) / 12.0 }; \
556
+ const double r_signsq { 0.5 / signsq };",
557
+ )
558
+
559
+
560
+ def gauss_spline(x, n):
561
+ r"""Gaussian approximation to B-spline basis function of order n.
562
+
563
+ Parameters
564
+ ----------
565
+ x : array_like
566
+ a knot vector
567
+ n : int
568
+ The order of the spline. Must be nonnegative, i.e. n >= 0
569
+
570
+ Returns
571
+ -------
572
+ res : ndarray
573
+ B-spline basis function values approximated by a zero-mean Gaussian
574
+ function.
575
+
576
+ Notes
577
+ -----
578
+ The B-spline basis function can be approximated well by a zero-mean
579
+ Gaussian function with standard-deviation equal to :math:`\sigma=(n+1)/12`
580
+ for large `n` :
581
+
582
+ .. math:: \frac{1}{\sqrt {2\pi\sigma^2}}exp(-\frac{x^2}{2\sigma})
583
+
584
+ See [1]_, [2]_ for more information.
585
+
586
+ References
587
+ ----------
588
+ .. [1] Bouma H., Vilanova A., Bescos J.O., ter Haar Romeny B.M., Gerritsen
589
+ F.A. (2007) Fast and Accurate Gaussian Derivatives Based on B-Splines.
590
+ In: Sgallari F., Murli A., Paragios N. (eds) Scale Space and Variational
591
+ Methods in Computer Vision. SSVM 2007. Lecture Notes in Computer
592
+ Science, vol 4485. Springer, Berlin, Heidelberg
593
+ .. [2] http://folk.uio.no/inf3330/scripting/doc/python/SciPy/tutorial/old/node24.html
594
+ """ # NOQA
595
+ x = cupy.asarray(x)
596
+ return _gauss_spline_kernel(x, n)
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_czt.py ADDED
@@ -0,0 +1,451 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This program is public domain
2
+ # Authors: Paul Kienzle, Nadav Horesh
3
+ #
4
+ # Adapted from scipy 1.10.1.
5
+ #
6
+ """
7
+ Chirp z-transform.
8
+
9
+ We provide two interfaces to the chirp z-transform: an object interface
10
+ which precalculates part of the transform and can be applied efficiently
11
+ to many different data sets, and a functional interface which is applied
12
+ only to the given data set.
13
+
14
+ Transforms
15
+ ----------
16
+
17
+ CZT : callable (x, axis=-1) -> array
18
+ Define a chirp z-transform that can be applied to different signals.
19
+ ZoomFFT : callable (x, axis=-1) -> array
20
+ Define a Fourier transform on a range of frequencies.
21
+
22
+ Functions
23
+ ---------
24
+
25
+ czt : array
26
+ Compute the chirp z-transform for a signal.
27
+ zoom_fft : array
28
+ Compute the Fourier transform on a range of frequencies.
29
+ """
30
+
31
+ import cmath
32
+ import numbers
33
+ import cupy
34
+ from numpy import pi
35
+ from cupyx.scipy.fft import fft, ifft, next_fast_len
36
+
37
+ __all__ = ['czt', 'zoom_fft', 'CZT', 'ZoomFFT', 'czt_points']
38
+
39
+
40
+ def _validate_sizes(n, m):
41
+ if n < 1 or not isinstance(n, numbers.Integral):
42
+ raise ValueError('Invalid number of CZT data '
43
+ f'points ({n}) specified. '
44
+ 'n must be positive and integer type.')
45
+
46
+ if m is None:
47
+ m = n
48
+ elif m < 1 or not isinstance(m, numbers.Integral):
49
+ raise ValueError('Invalid number of CZT output '
50
+ f'points ({m}) specified. '
51
+ 'm must be positive and integer type.')
52
+
53
+ return m
54
+
55
+
56
+ def czt_points(m, w=None, a=1+0j):
57
+ """
58
+ Return the points at which the chirp z-transform is computed.
59
+
60
+ Parameters
61
+ ----------
62
+ m : int
63
+ The number of points desired.
64
+ w : complex, optional
65
+ The ratio between points in each step.
66
+ Defaults to equally spaced points around the entire unit circle.
67
+ a : complex, optional
68
+ The starting point in the complex plane. Default is 1+0j.
69
+
70
+ Returns
71
+ -------
72
+ out : ndarray
73
+ The points in the Z plane at which `CZT` samples the z-transform,
74
+ when called with arguments `m`, `w`, and `a`, as complex numbers.
75
+
76
+ See Also
77
+ --------
78
+ CZT : Class that creates a callable chirp z-transform function.
79
+ czt : Convenience function for quickly calculating CZT.
80
+ scipy.signal.czt_points
81
+
82
+ """
83
+ m = _validate_sizes(1, m)
84
+
85
+ k = cupy.arange(m)
86
+
87
+ a = 1.0 * a # at least float
88
+
89
+ if w is None:
90
+ # Nothing specified, default to FFT
91
+ return a * cupy.exp(2j * pi * k / m)
92
+ else:
93
+ # w specified
94
+ w = 1.0 * w # at least float
95
+ return a * w**-k
96
+
97
+
98
+ class CZT:
99
+ """
100
+ Create a callable chirp z-transform function.
101
+
102
+ Transform to compute the frequency response around a spiral.
103
+ Objects of this class are callables which can compute the
104
+ chirp z-transform on their inputs. This object precalculates the constant
105
+ chirps used in the given transform.
106
+
107
+ Parameters
108
+ ----------
109
+ n : int
110
+ The size of the signal.
111
+ m : int, optional
112
+ The number of output points desired. Default is `n`.
113
+ w : complex, optional
114
+ The ratio between points in each step. This must be precise or the
115
+ accumulated error will degrade the tail of the output sequence.
116
+ Defaults to equally spaced points around the entire unit circle.
117
+ a : complex, optional
118
+ The starting point in the complex plane. Default is 1+0j.
119
+
120
+ Returns
121
+ -------
122
+ f : CZT
123
+ Callable object ``f(x, axis=-1)`` for computing the chirp z-transform
124
+ on `x`.
125
+
126
+ See Also
127
+ --------
128
+ czt : Convenience function for quickly calculating CZT.
129
+ ZoomFFT : Class that creates a callable partial FFT function.
130
+ scipy.signal.CZT
131
+
132
+ Notes
133
+ -----
134
+ The defaults are chosen such that ``f(x)`` is equivalent to
135
+ ``fft.fft(x)`` and, if ``m > len(x)``, that ``f(x, m)`` is equivalent to
136
+ ``fft.fft(x, m)``.
137
+
138
+ If `w` does not lie on the unit circle, then the transform will be
139
+ around a spiral with exponentially-increasing radius. Regardless,
140
+ angle will increase linearly.
141
+
142
+ For transforms that do lie on the unit circle, accuracy is better when
143
+ using `ZoomFFT`, since any numerical error in `w` is
144
+ accumulated for long data lengths, drifting away from the unit circle.
145
+
146
+ The chirp z-transform can be faster than an equivalent FFT with
147
+ zero padding. Try it with your own array sizes to see.
148
+
149
+ However, the chirp z-transform is considerably less precise than the
150
+ equivalent zero-padded FFT.
151
+
152
+ As this CZT is implemented using the Bluestein algorithm [1]_, it can
153
+ compute large prime-length Fourier transforms in O(N log N) time, rather
154
+ than the O(N**2) time required by the direct DFT calculation.
155
+ (`scipy.fft` also uses Bluestein's algorithm'.)
156
+
157
+ (The name "chirp z-transform" comes from the use of a chirp in the
158
+ Bluestein algorithm [2]_. It does not decompose signals into chirps, like
159
+ other transforms with "chirp" in the name.)
160
+
161
+ References
162
+ ----------
163
+ .. [1] Leo I. Bluestein, "A linear filtering approach to the computation
164
+ of the discrete Fourier transform," Northeast Electronics Research
165
+ and Engineering Meeting Record 10, 218-219 (1968).
166
+ .. [2] Rabiner, Schafer, and Rader, "The chirp z-transform algorithm and
167
+ its application," Bell Syst. Tech. J. 48, 1249-1292 (1969).
168
+
169
+ """
170
+
171
+ def __init__(self, n, m=None, w=None, a=1+0j):
172
+ m = _validate_sizes(n, m)
173
+
174
+ k = cupy.arange(max(m, n), dtype=cupy.min_scalar_type(-max(m, n)**2))
175
+
176
+ if w is None:
177
+ # Nothing specified, default to FFT-like
178
+ w = cmath.exp(-2j*pi/m)
179
+ wk2 = cupy.exp(-(1j * pi * ((k**2) % (2*m))) / m)
180
+ else:
181
+ # w specified
182
+ wk2 = w**(k**2/2.)
183
+
184
+ a = 1.0 * a # at least float
185
+
186
+ self.w, self.a = w, a
187
+ self.m, self.n = m, n
188
+
189
+ nfft = next_fast_len(n + m - 1)
190
+ self._Awk2 = a**-k[:n] * wk2[:n]
191
+ self._nfft = nfft
192
+ self._Fwk2 = fft(1/cupy.hstack((wk2[n-1:0:-1], wk2[:m])), nfft)
193
+ self._wk2 = wk2[:m]
194
+ self._yidx = slice(n-1, n+m-1)
195
+
196
+ def __call__(self, x, *, axis=-1):
197
+ """
198
+ Calculate the chirp z-transform of a signal.
199
+
200
+ Parameters
201
+ ----------
202
+ x : array
203
+ The signal to transform.
204
+ axis : int, optional
205
+ Axis over which to compute the FFT. If not given, the last axis is
206
+ used.
207
+
208
+ Returns
209
+ -------
210
+ out : ndarray
211
+ An array of the same dimensions as `x`, but with the length of the
212
+ transformed axis set to `m`.
213
+ """
214
+ x = cupy.asarray(x)
215
+ if x.shape[axis] != self.n:
216
+ raise ValueError(f"CZT defined for length {self.n}, not "
217
+ f"{x.shape[axis]}")
218
+ # Calculate transpose coordinates, to allow operation on any given axis
219
+ trnsp = list(range(x.ndim))
220
+ trnsp[axis], trnsp[-1] = trnsp[-1], trnsp[axis]
221
+ x = x.transpose(*trnsp)
222
+ y = ifft(self._Fwk2 * fft(x*self._Awk2, self._nfft))
223
+ y = y[..., self._yidx] * self._wk2
224
+ return y.transpose(*trnsp)
225
+
226
+ def points(self):
227
+ """
228
+ Return the points at which the chirp z-transform is computed.
229
+ """
230
+ return czt_points(self.m, self.w, self.a)
231
+
232
+
233
+ class ZoomFFT(CZT):
234
+ """
235
+ Create a callable zoom FFT transform function.
236
+
237
+ This is a specialization of the chirp z-transform (`CZT`) for a set of
238
+ equally-spaced frequencies around the unit circle, used to calculate a
239
+ section of the FFT more efficiently than calculating the entire FFT and
240
+ truncating. [1]_
241
+
242
+ Parameters
243
+ ----------
244
+ n : int
245
+ The size of the signal.
246
+ fn : array_like
247
+ A length-2 sequence [`f1`, `f2`] giving the frequency range, or a
248
+ scalar, for which the range [0, `fn`] is assumed.
249
+ m : int, optional
250
+ The number of points to evaluate. Default is `n`.
251
+ fs : float, optional
252
+ The sampling frequency. If ``fs=10`` represented 10 kHz, for example,
253
+ then `f1` and `f2` would also be given in kHz.
254
+ The default sampling frequency is 2, so `f1` and `f2` should be
255
+ in the range [0, 1] to keep the transform below the Nyquist
256
+ frequency.
257
+ endpoint : bool, optional
258
+ If True, `f2` is the last sample. Otherwise, it is not included.
259
+ Default is False.
260
+
261
+ Returns
262
+ -------
263
+ f : ZoomFFT
264
+ Callable object ``f(x, axis=-1)`` for computing the zoom FFT on `x`.
265
+
266
+ See Also
267
+ --------
268
+ zoom_fft : Convenience function for calculating a zoom FFT.
269
+ scipy.signal.ZoomFFT
270
+
271
+ Notes
272
+ -----
273
+ The defaults are chosen such that ``f(x, 2)`` is equivalent to
274
+ ``fft.fft(x)`` and, if ``m > len(x)``, that ``f(x, 2, m)`` is equivalent to
275
+ ``fft.fft(x, m)``.
276
+
277
+ Sampling frequency is 1/dt, the time step between samples in the
278
+ signal `x`. The unit circle corresponds to frequencies from 0 up
279
+ to the sampling frequency. The default sampling frequency of 2
280
+ means that `f1`, `f2` values up to the Nyquist frequency are in the
281
+ range [0, 1). For `f1`, `f2` values expressed in radians, a sampling
282
+ frequency of 2*pi should be used.
283
+
284
+ Remember that a zoom FFT can only interpolate the points of the existing
285
+ FFT. It cannot help to resolve two separate nearby frequencies.
286
+ Frequency resolution can only be increased by increasing acquisition
287
+ time.
288
+
289
+ These functions are implemented using Bluestein's algorithm (as is
290
+ `scipy.fft`). [2]_
291
+
292
+ References
293
+ ----------
294
+ .. [1] Steve Alan Shilling, "A study of the chirp z-transform and its
295
+ applications", pg 29 (1970)
296
+ https://krex.k-state.edu/dspace/bitstream/handle/2097/7844/LD2668R41972S43.pdf
297
+ .. [2] Leo I. Bluestein, "A linear filtering approach to the computation
298
+ of the discrete Fourier transform," Northeast Electronics Research
299
+ and Engineering Meeting Record 10, 218-219 (1968).
300
+ """
301
+
302
+ def __init__(self, n, fn, m=None, *, fs=2, endpoint=False):
303
+ m = _validate_sizes(n, m)
304
+
305
+ k = cupy.arange(max(m, n), dtype=cupy.min_scalar_type(-max(m, n)**2))
306
+
307
+ fn = cupy.asarray(fn)
308
+ if cupy.size(fn) == 2:
309
+ f1, f2 = fn
310
+ elif cupy.size(fn) == 1:
311
+ f1, f2 = 0.0, fn
312
+ else:
313
+ raise ValueError('fn must be a scalar or 2-length sequence')
314
+
315
+ self.f1, self.f2, self.fs = f1, f2, fs
316
+
317
+ if endpoint:
318
+ scale = ((f2 - f1) * m) / (fs * (m - 1))
319
+ else:
320
+ scale = (f2 - f1) / fs
321
+ a = cmath.exp(2j * pi * f1/fs)
322
+ wk2 = cupy.exp(-(1j * pi * scale * k**2) / m)
323
+
324
+ self.w = cmath.exp(-2j*pi/m * scale)
325
+ self.a = a
326
+ self.m, self.n = m, n
327
+
328
+ ak = cupy.exp(-2j * pi * f1/fs * k[:n])
329
+ self._Awk2 = ak * wk2[:n]
330
+
331
+ nfft = next_fast_len(n + m - 1)
332
+ self._nfft = nfft
333
+ self._Fwk2 = fft(1/cupy.hstack((wk2[n-1:0:-1], wk2[:m])), nfft)
334
+ self._wk2 = wk2[:m]
335
+ self._yidx = slice(n-1, n+m-1)
336
+
337
+
338
+ def czt(x, m=None, w=None, a=1+0j, *, axis=-1):
339
+ """
340
+ Compute the frequency response around a spiral in the Z plane.
341
+
342
+ Parameters
343
+ ----------
344
+ x : array
345
+ The signal to transform.
346
+ m : int, optional
347
+ The number of output points desired. Default is the length of the
348
+ input data.
349
+ w : complex, optional
350
+ The ratio between points in each step. This must be precise or the
351
+ accumulated error will degrade the tail of the output sequence.
352
+ Defaults to equally spaced points around the entire unit circle.
353
+ a : complex, optional
354
+ The starting point in the complex plane. Default is 1+0j.
355
+ axis : int, optional
356
+ Axis over which to compute the FFT. If not given, the last axis is
357
+ used.
358
+
359
+ Returns
360
+ -------
361
+ out : ndarray
362
+ An array of the same dimensions as `x`, but with the length of the
363
+ transformed axis set to `m`.
364
+
365
+ See Also
366
+ --------
367
+ CZT : Class that creates a callable chirp z-transform function.
368
+ zoom_fft : Convenience function for partial FFT calculations.
369
+ scipy.signal.czt
370
+
371
+ Notes
372
+ -----
373
+ The defaults are chosen such that ``signal.czt(x)`` is equivalent to
374
+ ``fft.fft(x)`` and, if ``m > len(x)``, that ``signal.czt(x, m)`` is
375
+ equivalent to ``fft.fft(x, m)``.
376
+
377
+ If the transform needs to be repeated, use `CZT` to construct a
378
+ specialized transform function which can be reused without
379
+ recomputing constants.
380
+
381
+ An example application is in system identification, repeatedly evaluating
382
+ small slices of the z-transform of a system, around where a pole is
383
+ expected to exist, to refine the estimate of the pole's true location. [1]_
384
+
385
+ References
386
+ ----------
387
+ .. [1] Steve Alan Shilling, "A study of the chirp z-transform and its
388
+ applications", pg 20 (1970)
389
+ https://krex.k-state.edu/dspace/bitstream/handle/2097/7844/LD2668R41972S43.pdf
390
+
391
+ """
392
+ x = cupy.asarray(x)
393
+ transform = CZT(x.shape[axis], m=m, w=w, a=a)
394
+ return transform(x, axis=axis)
395
+
396
+
397
+ def zoom_fft(x, fn, m=None, *, fs=2, endpoint=False, axis=-1):
398
+ """
399
+ Compute the DFT of `x` only for frequencies in range `fn`.
400
+
401
+ Parameters
402
+ ----------
403
+ x : array
404
+ The signal to transform.
405
+ fn : array_like
406
+ A length-2 sequence [`f1`, `f2`] giving the frequency range, or a
407
+ scalar, for which the range [0, `fn`] is assumed.
408
+ m : int, optional
409
+ The number of points to evaluate. The default is the length of `x`.
410
+ fs : float, optional
411
+ The sampling frequency. If ``fs=10`` represented 10 kHz, for example,
412
+ then `f1` and `f2` would also be given in kHz.
413
+ The default sampling frequency is 2, so `f1` and `f2` should be
414
+ in the range [0, 1] to keep the transform below the Nyquist
415
+ frequency.
416
+ endpoint : bool, optional
417
+ If True, `f2` is the last sample. Otherwise, it is not included.
418
+ Default is False.
419
+ axis : int, optional
420
+ Axis over which to compute the FFT. If not given, the last axis is
421
+ used.
422
+
423
+ Returns
424
+ -------
425
+ out : ndarray
426
+ The transformed signal. The Fourier transform will be calculated
427
+ at the points f1, f1+df, f1+2df, ..., f2, where df=(f2-f1)/m.
428
+
429
+ See Also
430
+ --------
431
+ ZoomFFT : Class that creates a callable partial FFT function.
432
+ scipy.signal.zoom_fft
433
+
434
+ Notes
435
+ -----
436
+ The defaults are chosen such that ``signal.zoom_fft(x, 2)`` is equivalent
437
+ to ``fft.fft(x)`` and, if ``m > len(x)``, that ``signal.zoom_fft(x, 2, m)``
438
+ is equivalent to ``fft.fft(x, m)``.
439
+
440
+ To graph the magnitude of the resulting transform, use::
441
+
442
+ plot(linspace(f1, f2, m, endpoint=False),
443
+ abs(zoom_fft(x, [f1, f2], m)))
444
+
445
+ If the transform needs to be repeated, use `ZoomFFT` to construct
446
+ a specialized transform function which can be reused without
447
+ recomputing constants.
448
+ """
449
+ x = cupy.asarray(x)
450
+ transform = ZoomFFT(x.shape[axis], fn, m=m, fs=fs, endpoint=endpoint)
451
+ return transform(x, axis=axis)
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_filter_design.py ADDED
@@ -0,0 +1,811 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import operator
2
+ from math import pi
3
+ import warnings
4
+
5
+ import cupy
6
+ from cupy.polynomial.polynomial import (
7
+ polyval as npp_polyval, polyvalfromroots as npp_polyvalfromroots)
8
+ import cupyx.scipy.fft as sp_fft
9
+ from cupyx import jit
10
+ from cupyx.scipy._lib._util import float_factorial
11
+ from cupyx.scipy.signal._polyutils import roots
12
+
13
+ EPSILON = 2e-16
14
+
15
+
16
+ def _try_convert_to_int(x):
17
+ """Return an integer for ``5`` and ``array(5)``, fail if not an
18
+ integer scalar.
19
+
20
+ NB: would be easier if ``operator.index(cupy.array(5))`` worked
21
+ (numpy.array(5) does)
22
+ """
23
+ if isinstance(x, cupy.ndarray):
24
+ if x.ndim == 0:
25
+ value = x.item()
26
+ else:
27
+ return x, False
28
+ else:
29
+ value = x
30
+ try:
31
+ return operator.index(value), True
32
+ except TypeError:
33
+ return value, False
34
+
35
+
36
+ def findfreqs(num, den, N, kind='ba'):
37
+ """
38
+ Find array of frequencies for computing the response of an analog filter.
39
+
40
+ Parameters
41
+ ----------
42
+ num, den : array_like, 1-D
43
+ The polynomial coefficients of the numerator and denominator of the
44
+ transfer function of the filter or LTI system, where the coefficients
45
+ are ordered from highest to lowest degree. Or, the roots of the
46
+ transfer function numerator and denominator (i.e., zeroes and poles).
47
+ N : int
48
+ The length of the array to be computed.
49
+ kind : str {'ba', 'zp'}, optional
50
+ Specifies whether the numerator and denominator are specified by their
51
+ polynomial coefficients ('ba'), or their roots ('zp').
52
+
53
+ Returns
54
+ -------
55
+ w : (N,) ndarray
56
+ A 1-D array of frequencies, logarithmically spaced.
57
+
58
+ Warning
59
+ -------
60
+ This function may synchronize the device.
61
+
62
+ See Also
63
+ --------
64
+ scipy.signal.find_freqs
65
+
66
+ Examples
67
+ --------
68
+ Find a set of nine frequencies that span the "interesting part" of the
69
+ frequency response for the filter with the transfer function
70
+
71
+ H(s) = s / (s^2 + 8s + 25)
72
+
73
+ >>> from scipy import signal
74
+ >>> signal.findfreqs([1, 0], [1, 8, 25], N=9)
75
+ array([ 1.00000000e-02, 3.16227766e-02, 1.00000000e-01,
76
+ 3.16227766e-01, 1.00000000e+00, 3.16227766e+00,
77
+ 1.00000000e+01, 3.16227766e+01, 1.00000000e+02])
78
+ """
79
+ if kind == 'ba':
80
+ ep = cupy.atleast_1d(roots(den)) + 0j
81
+ tz = cupy.atleast_1d(roots(num)) + 0j
82
+ elif kind == 'zp':
83
+ ep = cupy.atleast_1d(den) + 0j
84
+ tz = cupy.atleast_1d(num) + 0j
85
+ else:
86
+ raise ValueError("input must be one of {'ba', 'zp'}")
87
+
88
+ if len(ep) == 0:
89
+ ep = cupy.atleast_1d(-1000) + 0j
90
+
91
+ ez = cupy.r_[
92
+ cupy.compress(ep.imag >= 0, ep, axis=-1),
93
+ cupy.compress((abs(tz) < 1e5) & (tz.imag >= 0), tz, axis=-1)]
94
+
95
+ integ = cupy.abs(ez) < 1e-10
96
+ hfreq = cupy.around(cupy.log10(cupy.max(3 * cupy.abs(ez.real + integ) +
97
+ 1.5 * ez.imag)) + 0.5)
98
+ lfreq = cupy.around(cupy.log10(0.1 * cupy.min(cupy.abs((ez + integ).real) +
99
+ 2 * ez.imag)) - 0.5)
100
+ w = cupy.logspace(lfreq, hfreq, N)
101
+ return w
102
+
103
+
104
+ def freqs(b, a, worN=200, plot=None):
105
+ """
106
+ Compute frequency response of analog filter.
107
+
108
+ Given the M-order numerator `b` and N-order denominator `a` of an analog
109
+ filter, compute its frequency response::
110
+
111
+ b[0]*(jw)**M + b[1]*(jw)**(M-1) + ... + b[M]
112
+ H(w) = ----------------------------------------------
113
+ a[0]*(jw)**N + a[1]*(jw)**(N-1) + ... + a[N]
114
+
115
+ Parameters
116
+ ----------
117
+ b : array_like
118
+ Numerator of a linear filter.
119
+ a : array_like
120
+ Denominator of a linear filter.
121
+ worN : {None, int, array_like}, optional
122
+ If None, then compute at 200 frequencies around the interesting parts
123
+ of the response curve (determined by pole-zero locations). If a single
124
+ integer, then compute at that many frequencies. Otherwise, compute the
125
+ response at the angular frequencies (e.g., rad/s) given in `worN`.
126
+ plot : callable, optional
127
+ A callable that takes two arguments. If given, the return parameters
128
+ `w` and `h` are passed to plot. Useful for plotting the frequency
129
+ response inside `freqs`.
130
+
131
+ Returns
132
+ -------
133
+ w : ndarray
134
+ The angular frequencies at which `h` was computed.
135
+ h : ndarray
136
+ The frequency response.
137
+
138
+ See Also
139
+ --------
140
+ scipy.signal.freqs
141
+ freqz : Compute the frequency response of a digital filter.
142
+
143
+ """
144
+ if worN is None:
145
+ # For backwards compatibility
146
+ w = findfreqs(b, a, 200)
147
+
148
+ else:
149
+ N, _is_int = _try_convert_to_int(worN)
150
+ if _is_int:
151
+ w = findfreqs(b, a, N)
152
+ else:
153
+ w = cupy.atleast_1d(worN)
154
+
155
+ s = 1j * w
156
+ h = cupy.polyval(b, s) / cupy.polyval(a, s)
157
+
158
+ if plot is not None:
159
+ plot(w, h)
160
+
161
+ return w, h
162
+
163
+
164
+ def freqs_zpk(z, p, k, worN=200):
165
+ """
166
+ Compute frequency response of analog filter.
167
+
168
+ Given the zeros `z`, poles `p`, and gain `k` of a filter, compute its
169
+ frequency response::
170
+
171
+ (jw-z[0]) * (jw-z[1]) * ... * (jw-z[-1])
172
+ H(w) = k * ----------------------------------------
173
+ (jw-p[0]) * (jw-p[1]) * ... * (jw-p[-1])
174
+
175
+ Parameters
176
+ ----------
177
+ z : array_like
178
+ Zeroes of a linear filter
179
+ p : array_like
180
+ Poles of a linear filter
181
+ k : scalar
182
+ Gain of a linear filter
183
+ worN : {None, int, array_like}, optional
184
+ If None, then compute at 200 frequencies around the interesting parts
185
+ of the response curve (determined by pole-zero locations). If a single
186
+ integer, then compute at that many frequencies. Otherwise, compute the
187
+ response at the angular frequencies (e.g., rad/s) given in `worN`.
188
+
189
+ Returns
190
+ -------
191
+ w : ndarray
192
+ The angular frequencies at which `h` was computed.
193
+ h : ndarray
194
+ The frequency response.
195
+
196
+ See Also
197
+ --------
198
+ scipy.signal.freqs_zpk
199
+
200
+ """
201
+ k = cupy.asarray(k)
202
+ if k.size > 1:
203
+ raise ValueError('k must be a single scalar gain')
204
+
205
+ if worN is None:
206
+ # For backwards compatibility
207
+ w = findfreqs(z, p, 200, kind='zp')
208
+ else:
209
+ N, _is_int = _try_convert_to_int(worN)
210
+ if _is_int:
211
+ w = findfreqs(z, p, worN, kind='zp')
212
+ else:
213
+ w = worN
214
+
215
+ w = cupy.atleast_1d(w)
216
+ s = 1j * w
217
+ num = npp_polyvalfromroots(s, z)
218
+ den = npp_polyvalfromroots(s, p)
219
+ h = k * num/den
220
+ return w, h
221
+
222
+
223
+ def _is_int_type(x):
224
+ """
225
+ Check if input is of a scalar integer type (so ``5`` and ``array(5)`` will
226
+ pass, while ``5.0`` and ``array([5])`` will fail.
227
+ """
228
+ if cupy.ndim(x) != 0:
229
+ # Older versions of NumPy did not raise for np.array([1]).__index__()
230
+ # This is safe to remove when support for those versions is dropped
231
+ return False
232
+ try:
233
+ operator.index(x)
234
+ except TypeError:
235
+ return False
236
+ else:
237
+ return True
238
+
239
+
240
+ def group_delay(system, w=512, whole=False, fs=2 * cupy.pi):
241
+ r"""Compute the group delay of a digital filter.
242
+
243
+ The group delay measures by how many samples amplitude envelopes of
244
+ various spectral components of a signal are delayed by a filter.
245
+ It is formally defined as the derivative of continuous (unwrapped) phase::
246
+
247
+ d jw
248
+ D(w) = - -- arg H(e)
249
+ dw
250
+
251
+ Parameters
252
+ ----------
253
+ system : tuple of array_like (b, a)
254
+ Numerator and denominator coefficients of a filter transfer function.
255
+ w : {None, int, array_like}, optional
256
+ If a single integer, then compute at that many frequencies (default is
257
+ N=512).
258
+
259
+ If an array_like, compute the delay at the frequencies given. These
260
+ are in the same units as `fs`.
261
+ whole : bool, optional
262
+ Normally, frequencies are computed from 0 to the Nyquist frequency,
263
+ fs/2 (upper-half of unit-circle). If `whole` is True, compute
264
+ frequencies from 0 to fs. Ignored if w is array_like.
265
+ fs : float, optional
266
+ The sampling frequency of the digital system. Defaults to 2*pi
267
+ radians/sample (so w is from 0 to pi).
268
+
269
+ Returns
270
+ -------
271
+ w : ndarray
272
+ The frequencies at which group delay was computed, in the same units
273
+ as `fs`. By default, `w` is normalized to the range [0, pi)
274
+ (radians/sample).
275
+ gd : ndarray
276
+ The group delay.
277
+
278
+ See Also
279
+ --------
280
+ freqz : Frequency response of a digital filter
281
+
282
+ Notes
283
+ -----
284
+ The similar function in MATLAB is called `grpdelay`.
285
+
286
+ If the transfer function :math:`H(z)` has zeros or poles on the unit
287
+ circle, the group delay at corresponding frequencies is undefined.
288
+ When such a case arises the warning is raised and the group delay
289
+ is set to 0 at those frequencies.
290
+
291
+ For the details of numerical computation of the group delay refer to [1]_.
292
+
293
+ References
294
+ ----------
295
+ .. [1] Richard G. Lyons, "Understanding Digital Signal Processing,
296
+ 3rd edition", p. 830.
297
+
298
+ """
299
+ if w is None:
300
+ # For backwards compatibility
301
+ w = 512
302
+
303
+ if _is_int_type(w):
304
+ if whole:
305
+ w = cupy.linspace(0, 2 * cupy.pi, w, endpoint=False)
306
+ else:
307
+ w = cupy.linspace(0, cupy.pi, w, endpoint=False)
308
+ else:
309
+ w = cupy.atleast_1d(w)
310
+ w = 2 * cupy.pi * w / fs
311
+
312
+ b, a = map(cupy.atleast_1d, system)
313
+ c = cupy.convolve(b, a[::-1])
314
+ cr = c * cupy.arange(c.size)
315
+ z = cupy.exp(-1j * w)
316
+ num = cupy.polyval(cr[::-1], z)
317
+ den = cupy.polyval(c[::-1], z)
318
+ gd = cupy.real(num / den) - a.size + 1
319
+ singular = ~cupy.isfinite(gd)
320
+ gd[singular] = 0
321
+
322
+ w = w * fs / (2 * cupy.pi)
323
+ return w, gd
324
+
325
+
326
+ def freqz(b, a=1, worN=512, whole=False, plot=None, fs=2*pi,
327
+ include_nyquist=False):
328
+ """
329
+ Compute the frequency response of a digital filter.
330
+
331
+ Given the M-order numerator `b` and N-order denominator `a` of a digital
332
+ filter, compute its frequency response::
333
+
334
+ jw -jw -jwM
335
+ jw B(e ) b[0] + b[1]e + ... + b[M]e
336
+ H(e ) = ------ = -----------------------------------
337
+ jw -jw -jwN
338
+ A(e ) a[0] + a[1]e + ... + a[N]e
339
+
340
+ Parameters
341
+ ----------
342
+ b : array_like
343
+ Numerator of a linear filter. If `b` has dimension greater than 1,
344
+ it is assumed that the coefficients are stored in the first dimension,
345
+ and ``b.shape[1:]``, ``a.shape[1:]``, and the shape of the frequencies
346
+ array must be compatible for broadcasting.
347
+ a : array_like
348
+ Denominator of a linear filter. If `b` has dimension greater than 1,
349
+ it is assumed that the coefficients are stored in the first dimension,
350
+ and ``b.shape[1:]``, ``a.shape[1:]``, and the shape of the frequencies
351
+ array must be compatible for broadcasting.
352
+ worN : {None, int, array_like}, optional
353
+ If a single integer, then compute at that many frequencies (default is
354
+ N=512). This is a convenient alternative to::
355
+
356
+ cupy.linspace(0, fs if whole else fs/2, N,
357
+ endpoint=include_nyquist)
358
+
359
+ Using a number that is fast for FFT computations can result in
360
+ faster computations (see Notes).
361
+
362
+ If an array_like, compute the response at the frequencies given.
363
+ These are in the same units as `fs`.
364
+ whole : bool, optional
365
+ Normally, frequencies are computed from 0 to the Nyquist frequency,
366
+ fs/2 (upper-half of unit-circle). If `whole` is True, compute
367
+ frequencies from 0 to fs. Ignored if worN is array_like.
368
+ plot : callable
369
+ A callable that takes two arguments. If given, the return parameters
370
+ `w` and `h` are passed to plot. Useful for plotting the frequency
371
+ response inside `freqz`.
372
+ fs : float, optional
373
+ The sampling frequency of the digital system. Defaults to 2*pi
374
+ radians/sample (so w is from 0 to pi).
375
+ include_nyquist : bool, optional
376
+ If `whole` is False and `worN` is an integer, setting `include_nyquist`
377
+ to True will include the last frequency (Nyquist frequency) and is
378
+ otherwise ignored.
379
+
380
+ Returns
381
+ -------
382
+ w : ndarray
383
+ The frequencies at which `h` was computed, in the same units as `fs`.
384
+ By default, `w` is normalized to the range [0, pi) (radians/sample).
385
+ h : ndarray
386
+ The frequency response, as complex numbers.
387
+
388
+ See Also
389
+ --------
390
+ freqz_zpk
391
+ sosfreqz
392
+ scipy.signal.freqz
393
+
394
+
395
+ Notes
396
+ -----
397
+ Using Matplotlib's :func:`matplotlib.pyplot.plot` function as the callable
398
+ for `plot` produces unexpected results, as this plots the real part of the
399
+ complex transfer function, not the magnitude.
400
+ Try ``lambda w, h: plot(w, cupy.abs(h))``.
401
+
402
+ A direct computation via (R)FFT is used to compute the frequency response
403
+ when the following conditions are met:
404
+
405
+ 1. An integer value is given for `worN`.
406
+ 2. `worN` is fast to compute via FFT (i.e.,
407
+ `next_fast_len(worN) <scipy.fft.next_fast_len>` equals `worN`).
408
+ 3. The denominator coefficients are a single value (``a.shape[0] == 1``).
409
+ 4. `worN` is at least as long as the numerator coefficients
410
+ (``worN >= b.shape[0]``).
411
+ 5. If ``b.ndim > 1``, then ``b.shape[-1] == 1``.
412
+
413
+ For long FIR filters, the FFT approach can have lower error and be much
414
+ faster than the equivalent direct polynomial calculation.
415
+ """
416
+ b = cupy.atleast_1d(b)
417
+ a = cupy.atleast_1d(a)
418
+
419
+ if worN is None:
420
+ # For backwards compatibility
421
+ worN = 512
422
+
423
+ h = None
424
+
425
+ N, _is_int = _try_convert_to_int(worN)
426
+ if _is_int:
427
+ if N < 0:
428
+ raise ValueError(f'worN must be nonnegative, got {N}')
429
+ lastpoint = 2 * pi if whole else pi
430
+
431
+ # if include_nyquist is true and whole is false, w should
432
+ # include end point
433
+ w = cupy.linspace(
434
+ 0, lastpoint, N, endpoint=include_nyquist and not whole)
435
+
436
+ use_fft = (a.size == 1 and
437
+ N >= b.shape[0] and
438
+ sp_fft.next_fast_len(N) == N and
439
+ (b.ndim == 1 or (b.shape[-1] == 1))
440
+ )
441
+
442
+ if use_fft:
443
+ # if N is fast, 2 * N will be fast, too, so no need to check
444
+ n_fft = N if whole else N * 2
445
+ if cupy.isrealobj(b) and cupy.isrealobj(a):
446
+ fft_func = sp_fft.rfft
447
+ else:
448
+ fft_func = sp_fft.fft
449
+
450
+ h = fft_func(b, n=n_fft, axis=0)[:N]
451
+ h /= a
452
+ if fft_func is sp_fft.rfft and whole:
453
+ # exclude DC and maybe Nyquist (no need to use axis_reverse
454
+ # here because we can build reversal with the truncation)
455
+ stop = -1 if n_fft % 2 == 1 else -2
456
+ h_flip = slice(stop, 0, -1)
457
+ h = cupy.concatenate((h, h[h_flip].conj()))
458
+ if b.ndim > 1:
459
+ # Last axis of h has length 1, so drop it.
460
+ h = h[..., 0]
461
+ # Move the first axis of h to the end.
462
+ h = cupy.moveaxis(h, 0, -1)
463
+ else:
464
+ w = cupy.atleast_1d(worN)
465
+ w = 2 * pi * w / fs
466
+
467
+ if h is None: # still need to compute using freqs w
468
+ zm1 = cupy.exp(-1j * w)
469
+ h = (npp_polyval(zm1, b, tensor=False) /
470
+ npp_polyval(zm1, a, tensor=False))
471
+
472
+ w = w * fs / (2 * pi)
473
+
474
+ if plot is not None:
475
+ plot(w, h)
476
+
477
+ return w, h
478
+
479
+
480
+ def freqz_zpk(z, p, k, worN=512, whole=False, fs=2*pi):
481
+ r"""
482
+ Compute the frequency response of a digital filter in ZPK form.
483
+
484
+ Given the Zeros, Poles and Gain of a digital filter, compute its frequency
485
+ response:
486
+
487
+ :math:`H(z)=k \prod_i (z - Z[i]) / \prod_j (z - P[j])`
488
+
489
+ where :math:`k` is the `gain`, :math:`Z` are the `zeros` and :math:`P` are
490
+ the `poles`.
491
+
492
+ Parameters
493
+ ----------
494
+ z : array_like
495
+ Zeroes of a linear filter
496
+ p : array_like
497
+ Poles of a linear filter
498
+ k : scalar
499
+ Gain of a linear filter
500
+ worN : {None, int, array_like}, optional
501
+ If a single integer, then compute at that many frequencies (default is
502
+ N=512).
503
+
504
+ If an array_like, compute the response at the frequencies given.
505
+ These are in the same units as `fs`.
506
+ whole : bool, optional
507
+ Normally, frequencies are computed from 0 to the Nyquist frequency,
508
+ fs/2 (upper-half of unit-circle). If `whole` is True, compute
509
+ frequencies from 0 to fs. Ignored if w is array_like.
510
+ fs : float, optional
511
+ The sampling frequency of the digital system. Defaults to 2*pi
512
+ radians/sample (so w is from 0 to pi).
513
+
514
+ Returns
515
+ -------
516
+ w : ndarray
517
+ The frequencies at which `h` was computed, in the same units as `fs`.
518
+ By default, `w` is normalized to the range [0, pi) (radians/sample).
519
+ h : ndarray
520
+ The frequency response, as complex numbers.
521
+
522
+ See Also
523
+ --------
524
+ freqs : Compute the frequency response of an analog filter in TF form
525
+ freqs_zpk : Compute the frequency response of an analog filter in ZPK form
526
+ freqz : Compute the frequency response of a digital filter in TF form
527
+ scipy.signal.freqz_zpk
528
+
529
+ """
530
+ z, p = map(cupy.atleast_1d, (z, p))
531
+
532
+ if whole:
533
+ lastpoint = 2 * pi
534
+ else:
535
+ lastpoint = pi
536
+
537
+ if worN is None:
538
+ # For backwards compatibility
539
+ w = cupy.linspace(0, lastpoint, 512, endpoint=False)
540
+ else:
541
+ N, _is_int = _try_convert_to_int(worN)
542
+ if _is_int:
543
+ w = cupy.linspace(0, lastpoint, N, endpoint=False)
544
+ else:
545
+ w = cupy.atleast_1d(worN)
546
+ w = 2 * pi * w / fs
547
+
548
+ zm1 = cupy.exp(1j * w)
549
+ h = k * npp_polyvalfromroots(zm1, z) / npp_polyvalfromroots(zm1, p)
550
+
551
+ w = w * fs / (2 * pi)
552
+
553
+ return w, h
554
+
555
+
556
+ def _validate_sos(sos):
557
+ """Helper to validate a SOS input"""
558
+ sos = cupy.atleast_2d(sos)
559
+ if sos.ndim != 2:
560
+ raise ValueError('sos array must be 2D')
561
+ n_sections, m = sos.shape
562
+ if m != 6:
563
+ raise ValueError('sos array must be shape (n_sections, 6)')
564
+ if ((sos[:, 3] - 1) > 1e-15).any():
565
+ raise ValueError('sos[:, 3] should be all ones')
566
+ return sos, n_sections
567
+
568
+
569
+ def sosfreqz(sos, worN=512, whole=False, fs=2*pi):
570
+ r"""
571
+ Compute the frequency response of a digital filter in SOS format.
572
+
573
+ Given `sos`, an array with shape (n, 6) of second order sections of
574
+ a digital filter, compute the frequency response of the system function::
575
+
576
+ B0(z) B1(z) B{n-1}(z)
577
+ H(z) = ----- * ----- * ... * ---------
578
+ A0(z) A1(z) A{n-1}(z)
579
+
580
+ for z = exp(omega*1j), where B{k}(z) and A{k}(z) are numerator and
581
+ denominator of the transfer function of the k-th second order section.
582
+
583
+ Parameters
584
+ ----------
585
+ sos : array_like
586
+ Array of second-order filter coefficients, must have shape
587
+ ``(n_sections, 6)``. Each row corresponds to a second-order
588
+ section, with the first three columns providing the numerator
589
+ coefficients and the last three providing the denominator
590
+ coefficients.
591
+ worN : {None, int, array_like}, optional
592
+ If a single integer, then compute at that many frequencies (default is
593
+ N=512). Using a number that is fast for FFT computations can result
594
+ in faster computations (see Notes of `freqz`).
595
+
596
+ If an array_like, compute the response at the frequencies given (must
597
+ be 1-D). These are in the same units as `fs`.
598
+ whole : bool, optional
599
+ Normally, frequencies are computed from 0 to the Nyquist frequency,
600
+ fs/2 (upper-half of unit-circle). If `whole` is True, compute
601
+ frequencies from 0 to fs.
602
+ fs : float, optional
603
+ The sampling frequency of the digital system. Defaults to 2*pi
604
+ radians/sample (so w is from 0 to pi).
605
+
606
+ .. versionadded:: 1.2.0
607
+
608
+ Returns
609
+ -------
610
+ w : ndarray
611
+ The frequencies at which `h` was computed, in the same units as `fs`.
612
+ By default, `w` is normalized to the range [0, pi) (radians/sample).
613
+ h : ndarray
614
+ The frequency response, as complex numbers.
615
+
616
+ See Also
617
+ --------
618
+ freqz, sosfilt
619
+ scipy.signal.sosfreqz
620
+ """
621
+ sos, n_sections = _validate_sos(sos)
622
+ if n_sections == 0:
623
+ raise ValueError('Cannot compute frequencies with no sections')
624
+ h = 1.
625
+ for row in sos:
626
+ w, rowh = freqz(row[:3], row[3:], worN=worN, whole=whole, fs=fs)
627
+ h *= rowh
628
+ return w, h
629
+
630
+
631
+ def _hz_to_erb(hz):
632
+ """
633
+ Utility for converting from frequency (Hz) to the
634
+ Equivalent Rectangular Bandwidth (ERB) scale
635
+ ERB = frequency / EarQ + minBW
636
+ """
637
+ EarQ = 9.26449
638
+ minBW = 24.7
639
+ return hz / EarQ + minBW
640
+
641
+
642
+ @jit.rawkernel()
643
+ def _gammatone_iir_kernel(fs, freq, b, a):
644
+ tid = jit.blockIdx.x * jit.blockDim.x + jit.threadIdx.x
645
+
646
+ EarQ = 9.26449
647
+ minBW = 24.7
648
+ erb = freq / EarQ + minBW
649
+
650
+ T = 1./fs
651
+ bw = 2 * cupy.pi * 1.019 * erb
652
+ fr = 2 * freq * cupy.pi * T
653
+ bwT = bw * T
654
+
655
+ # Calculate the gain to normalize the volume at the center frequency
656
+ g1 = -2 * cupy.exp(2j * fr) * T
657
+ g2 = 2 * cupy.exp(-(bwT) + 1j * fr) * T
658
+ g3 = cupy.sqrt(3 + 2 ** (3 / 2)) * cupy.sin(fr)
659
+ g4 = cupy.sqrt(3 - 2 ** (3 / 2)) * cupy.sin(fr)
660
+ g5 = cupy.exp(2j * fr)
661
+
662
+ g = g1 + g2 * (cupy.cos(fr) - g4)
663
+ g *= (g1 + g2 * (cupy.cos(fr) + g4))
664
+ g *= (g1 + g2 * (cupy.cos(fr) - g3))
665
+ g *= (g1 + g2 * (cupy.cos(fr) + g3))
666
+ g /= ((-2 / cupy.exp(2 * bwT) - 2 * g5 + 2 * (1 + g5) /
667
+ cupy.exp(bwT)) ** 4)
668
+ g_act = cupy.abs(g)
669
+
670
+ # Calculate the numerator coefficients
671
+ if tid == 0:
672
+ b[tid] = (T ** 4) / g_act
673
+ a[tid] = 1
674
+ elif tid == 1:
675
+ b[tid] = -4 * T ** 4 * cupy.cos(fr) / cupy.exp(bw * T) / g_act
676
+ a[tid] = -8 * cupy.cos(fr) / cupy.exp(bw * T)
677
+ elif tid == 2:
678
+ b[tid] = 6 * T ** 4 * cupy.cos(2 * fr) / cupy.exp(2 * bw * T) / g_act
679
+ a[tid] = 4 * (4 + 3 * cupy.cos(2 * fr)) / cupy.exp(2 * bw * T)
680
+ elif tid == 3:
681
+ b[tid] = -4 * T ** 4 * cupy.cos(3 * fr) / cupy.exp(3 * bw * T) / g_act
682
+ a[tid] = -8 * (6 * cupy.cos(fr) + cupy.cos(3 * fr))
683
+ a[tid] /= cupy.exp(3 * bw * T)
684
+ elif tid == 4:
685
+ b[tid] = T ** 4 * cupy.cos(4 * fr) / cupy.exp(4 * bw * T) / g_act
686
+ a[tid] = 2 * (18 + 16 * cupy.cos(2 * fr) + cupy.cos(4 * fr))
687
+ a[tid] /= cupy.exp(4 * bw * T)
688
+ elif tid == 5:
689
+ a[tid] = -8 * (6 * cupy.cos(fr) + cupy.cos(3 * fr))
690
+ a[tid] /= cupy.exp(5 * bw * T)
691
+ elif tid == 6:
692
+ a[tid] = 4 * (4 + 3 * cupy.cos(2 * fr)) / cupy.exp(6 * bw * T)
693
+ elif tid == 7:
694
+ a[tid] = -8 * cupy.cos(fr) / cupy.exp(7 * bw * T)
695
+ elif tid == 8:
696
+ a[tid] = cupy.exp(-8 * bw * T)
697
+
698
+
699
+ def gammatone(freq, ftype, order=None, numtaps=None, fs=None):
700
+ """
701
+ Gammatone filter design.
702
+
703
+ This function computes the coefficients of an FIR or IIR gammatone
704
+ digital filter [1]_.
705
+
706
+ Parameters
707
+ ----------
708
+ freq : float
709
+ Center frequency of the filter (expressed in the same units
710
+ as `fs`).
711
+ ftype : {'fir', 'iir'}
712
+ The type of filter the function generates. If 'fir', the function
713
+ will generate an Nth order FIR gammatone filter. If 'iir', the
714
+ function will generate an 8th order digital IIR filter, modeled as
715
+ as 4th order gammatone filter.
716
+ order : int, optional
717
+ The order of the filter. Only used when ``ftype='fir'``.
718
+ Default is 4 to model the human auditory system. Must be between
719
+ 0 and 24.
720
+ numtaps : int, optional
721
+ Length of the filter. Only used when ``ftype='fir'``.
722
+ Default is ``fs*0.015`` if `fs` is greater than 1000,
723
+ 15 if `fs` is less than or equal to 1000.
724
+ fs : float, optional
725
+ The sampling frequency of the signal. `freq` must be between
726
+ 0 and ``fs/2``. Default is 2.
727
+
728
+ Returns
729
+ -------
730
+ b, a : ndarray, ndarray
731
+ Numerator (``b``) and denominator (``a``) polynomials of the filter.
732
+
733
+ Raises
734
+ ------
735
+ ValueError
736
+ If `freq` is less than or equal to 0 or greater than or equal to
737
+ ``fs/2``, if `ftype` is not 'fir' or 'iir', if `order` is less than
738
+ or equal to 0 or greater than 24 when ``ftype='fir'``
739
+
740
+ See Also
741
+ --------
742
+ firwin
743
+ iirfilter
744
+
745
+ References
746
+ ----------
747
+ .. [1] Slaney, Malcolm, "An Efficient Implementation of the
748
+ Patterson-Holdsworth Auditory Filter Bank", Apple Computer
749
+ Technical Report 35, 1993, pp.3-8, 34-39.
750
+ """
751
+ # Converts freq to float
752
+ freq = float(freq)
753
+
754
+ # Set sampling rate if not passed
755
+ if fs is None:
756
+ fs = 2
757
+ fs = float(fs)
758
+
759
+ # Check for invalid cutoff frequency or filter type
760
+ ftype = ftype.lower()
761
+ filter_types = ['fir', 'iir']
762
+ if not 0 < freq < fs / 2:
763
+ raise ValueError("The frequency must be between 0 and {}"
764
+ " (nyquist), but given {}.".format(fs / 2, freq))
765
+ if ftype not in filter_types:
766
+ raise ValueError('ftype must be either fir or iir.')
767
+
768
+ # Calculate FIR gammatone filter
769
+ if ftype == 'fir':
770
+ # Set order and numtaps if not passed
771
+ if order is None:
772
+ order = 4
773
+ order = operator.index(order)
774
+
775
+ if numtaps is None:
776
+ numtaps = max(int(fs * 0.015), 15)
777
+ numtaps = operator.index(numtaps)
778
+
779
+ # Check for invalid order
780
+ if not 0 < order <= 24:
781
+ raise ValueError("Invalid order: order must be > 0 and <= 24.")
782
+
783
+ # Gammatone impulse response settings
784
+ t = cupy.arange(numtaps) / fs
785
+ bw = 1.019 * _hz_to_erb(freq)
786
+
787
+ # Calculate the FIR gammatone filter
788
+ b = (t ** (order - 1)) * cupy.exp(-2 * cupy.pi * bw * t)
789
+ b *= cupy.cos(2 * cupy.pi * freq * t)
790
+
791
+ # Scale the FIR filter so the frequency response is 1 at cutoff
792
+ scale_factor = 2 * (2 * cupy.pi * bw) ** (order)
793
+ scale_factor /= float_factorial(order - 1)
794
+ scale_factor /= fs
795
+ b *= scale_factor
796
+ a = [1.0]
797
+
798
+ # Calculate IIR gammatone filter
799
+ elif ftype == 'iir':
800
+ # Raise warning if order and/or numtaps is passed
801
+ if order is not None:
802
+ warnings.warn('order is not used for IIR gammatone filter.')
803
+ if numtaps is not None:
804
+ warnings.warn('numtaps is not used for IIR gammatone filter.')
805
+
806
+ # Create empty filter coefficient lists
807
+ b = cupy.empty(5)
808
+ a = cupy.empty(9)
809
+ _gammatone_iir_kernel((9,), (1,), (fs, freq, b, a))
810
+
811
+ return b, a
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_fir_filter_design.py ADDED
@@ -0,0 +1,918 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Functions for FIR filter design."""
2
+ import math
3
+
4
+ from cupy.fft import fft, ifft
5
+ from cupy.linalg import solve, lstsq, LinAlgError
6
+ from cupyx.scipy.linalg import toeplitz, hankel
7
+ import cupyx
8
+ from cupyx.scipy.signal.windows import get_window
9
+
10
+ import cupy
11
+ import numpy
12
+
13
+
14
+ __all__ = ["firls", "minimum_phase"]
15
+
16
+
17
+ def kaiser_beta(a):
18
+ """Compute the Kaiser parameter `beta`, given the attenuation `a`.
19
+
20
+ Parameters
21
+ ----------
22
+ a : float
23
+ The desired attenuation in the stopband and maximum ripple in
24
+ the passband, in dB. This should be a *positive* number.
25
+
26
+ Returns
27
+ -------
28
+ beta : float
29
+ The `beta` parameter to be used in the formula for a Kaiser window.
30
+
31
+ References
32
+ ----------
33
+ Oppenheim, Schafer, "Discrete-Time Signal Processing", p.475-476.
34
+
35
+ See Also
36
+ --------
37
+ scipy.signal.kaiser_beta
38
+
39
+ """
40
+ if a > 50:
41
+ beta = 0.1102 * (a - 8.7)
42
+ elif a > 21:
43
+ beta = 0.5842 * (a - 21) ** 0.4 + 0.07886 * (a - 21)
44
+ else:
45
+ beta = 0.0
46
+ return beta
47
+
48
+
49
+ def kaiser_atten(numtaps, width):
50
+ """Compute the attenuation of a Kaiser FIR filter.
51
+
52
+ Given the number of taps `N` and the transition width `width`, compute the
53
+ attenuation `a` in dB, given by Kaiser's formula:
54
+
55
+ a = 2.285 * (N - 1) * pi * width + 7.95
56
+
57
+ Parameters
58
+ ----------
59
+ numtaps : int
60
+ The number of taps in the FIR filter.
61
+ width : float
62
+ The desired width of the transition region between passband and
63
+ stopband (or, in general, at any discontinuity) for the filter,
64
+ expressed as a fraction of the Nyquist frequency.
65
+
66
+ Returns
67
+ -------
68
+ a : float
69
+ The attenuation of the ripple, in dB.
70
+
71
+ See Also
72
+ --------
73
+ scipy.signal.kaiser_atten
74
+ """
75
+ a = 2.285 * (numtaps - 1) * cupy.pi * width + 7.95
76
+ return a
77
+
78
+
79
+ def kaiserord(ripple, width):
80
+ """
81
+ Determine the filter window parameters for the Kaiser window method.
82
+
83
+ The parameters returned by this function are generally used to create
84
+ a finite impulse response filter using the window method, with either
85
+ `firwin` or `firwin2`.
86
+
87
+ Parameters
88
+ ----------
89
+ ripple : float
90
+ Upper bound for the deviation (in dB) of the magnitude of the
91
+ filter's frequency response from that of the desired filter (not
92
+ including frequencies in any transition intervals). That is, if w
93
+ is the frequency expressed as a fraction of the Nyquist frequency,
94
+ A(w) is the actual frequency response of the filter and D(w) is the
95
+ desired frequency response, the design requirement is that::
96
+
97
+ abs(A(w) - D(w))) < 10**(-ripple/20)
98
+
99
+ for 0 <= w <= 1 and w not in a transition interval.
100
+ width : float
101
+ Width of transition region, normalized so that 1 corresponds to pi
102
+ radians / sample. That is, the frequency is expressed as a fraction
103
+ of the Nyquist frequency.
104
+
105
+ Returns
106
+ -------
107
+ numtaps : int
108
+ The length of the Kaiser window.
109
+ beta : float
110
+ The beta parameter for the Kaiser window.
111
+
112
+ See Also
113
+ --------
114
+ scipy.signal.kaiserord
115
+
116
+
117
+ """
118
+ A = abs(ripple) # in case somebody is confused as to what's meant
119
+ if A < 8:
120
+ # Formula for N is not valid in this range.
121
+ raise ValueError("Requested maximum ripple attenuation %f is too "
122
+ "small for the Kaiser formula." % A)
123
+ beta = kaiser_beta(A)
124
+
125
+ # Kaiser's formula (as given in Oppenheim and Schafer) is for the filter
126
+ # order, so we have to add 1 to get the number of taps.
127
+ numtaps = (A - 7.95) / 2.285 / (cupy.pi * width) + 1
128
+
129
+ return int(numpy.ceil(numtaps)), beta
130
+
131
+
132
+ _firwin_kernel = cupy.ElementwiseKernel(
133
+ "float64 win, int32 numtaps, raw float64 bands, int32 steps, bool scale",
134
+ "float64 h, float64 hc",
135
+ """
136
+ const double m { static_cast<double>( i ) - alpha ?
137
+ static_cast<double>( i ) - alpha : 1.0e-20 };
138
+
139
+ double temp {};
140
+ double left {};
141
+ double right {};
142
+
143
+ for ( int s = 0; s < steps; s++ ) {
144
+ left = bands[s * 2 + 0] ? bands[s * 2 + 0] : 1.0e-20;
145
+ right = bands[s * 2 + 1] ? bands[s * 2 + 1] : 1.0e-20;
146
+
147
+ temp += right * ( sin( right * m * M_PI ) / ( right * m * M_PI ) );
148
+ temp -= left * ( sin( left * m * M_PI ) / ( left * m * M_PI ) );
149
+ }
150
+
151
+ temp *= win;
152
+ h = temp;
153
+
154
+ double scale_frequency {};
155
+
156
+ if ( scale ) {
157
+ left = bands[0];
158
+ right = bands[1];
159
+
160
+ if ( left == 0 ) {
161
+ scale_frequency = 0.0;
162
+ } else if ( right == 1 ) {
163
+ scale_frequency = 1.0;
164
+ } else {
165
+ scale_frequency = 0.5 * ( left + right );
166
+ }
167
+ double c { cos( M_PI * m * scale_frequency ) };
168
+ hc = temp * c;
169
+ }
170
+ """,
171
+ "_firwin_kernel",
172
+ options=("-std=c++11",),
173
+ loop_prep="const double alpha { 0.5 * ( numtaps - 1 ) };",
174
+ )
175
+
176
+
177
+ # Scipy <= 1.12 has a deprecated `nyq` argument (nyq = fs/2).
178
+ # Remove it here, to be forward-looking.
179
+ def firwin(
180
+ numtaps,
181
+ cutoff,
182
+ width=None,
183
+ window="hamming",
184
+ pass_zero=True,
185
+ scale=True,
186
+ fs=2,
187
+ ):
188
+ """
189
+ FIR filter design using the window method.
190
+
191
+ This function computes the coefficients of a finite impulse response
192
+ filter. The filter will have linear phase; it will be Type I if
193
+ `numtaps` is odd and Type II if `numtaps` is even.
194
+
195
+ Type II filters always have zero response at the Nyquist frequency, so a
196
+ ValueError exception is raised if firwin is called with `numtaps` even and
197
+ having a passband whose right end is at the Nyquist frequency.
198
+
199
+ Parameters
200
+ ----------
201
+ numtaps : int
202
+ Length of the filter (number of coefficients, i.e. the filter
203
+ order + 1). `numtaps` must be odd if a passband includes the
204
+ Nyquist frequency.
205
+ cutoff : float or 1D array_like
206
+ Cutoff frequency of filter (expressed in the same units as `fs`)
207
+ OR an array of cutoff frequencies (that is, band edges). In the
208
+ latter case, the frequencies in `cutoff` should be positive and
209
+ monotonically increasing between 0 and `fs/2`. The values 0 and
210
+ `fs/2` must not be included in `cutoff`.
211
+ width : float or None, optional
212
+ If `width` is not None, then assume it is the approximate width
213
+ of the transition region (expressed in the same units as `fs`)
214
+ for use in Kaiser FIR filter design. In this case, the `window`
215
+ argument is ignored.
216
+ window : string or tuple of string and parameter values, optional
217
+ Desired window to use. See `cusignal.get_window` for a list
218
+ of windows and required parameters.
219
+ pass_zero : {True, False, 'bandpass', 'lowpass', 'highpass', 'bandstop'},
220
+ optional
221
+ If True, the gain at the frequency 0 (i.e. the "DC gain") is 1.
222
+ If False, the DC gain is 0. Can also be a string argument for the
223
+ desired filter type (equivalent to ``btype`` in IIR design functions).
224
+ scale : bool, optional
225
+ Set to True to scale the coefficients so that the frequency
226
+ response is exactly unity at a certain frequency.
227
+ That frequency is either:
228
+
229
+ - 0 (DC) if the first passband starts at 0 (i.e. pass_zero
230
+ is True)
231
+ - `fs/2` (the Nyquist frequency) if the first passband ends at
232
+ `fs/2` (i.e the filter is a single band highpass filter);
233
+ center of first passband otherwise
234
+ fs : float, optional
235
+ The sampling frequency of the signal. Each frequency in `cutoff`
236
+ must be between 0 and ``fs/2``. Default is 2.
237
+
238
+ Returns
239
+ -------
240
+ h : (numtaps,) ndarray
241
+ Coefficients of length `numtaps` FIR filter.
242
+
243
+ Raises
244
+ ------
245
+ ValueError
246
+ If any value in `cutoff` is less than or equal to 0 or greater
247
+ than or equal to ``fs/2``, if the values in `cutoff` are not strictly
248
+ monotonically increasing, or if `numtaps` is even but a passband
249
+ includes the Nyquist frequency.
250
+
251
+ See Also
252
+ --------
253
+ firwin2
254
+ firls
255
+ minimum_phase
256
+ remez
257
+
258
+ Examples
259
+ --------
260
+ Low-pass from 0 to f:
261
+
262
+ >>> import cusignal
263
+ >>> numtaps = 3
264
+ >>> f = 0.1
265
+ >>> cusignal.firwin(numtaps, f)
266
+ array([ 0.06799017, 0.86401967, 0.06799017])
267
+
268
+ Use a specific window function:
269
+
270
+ >>> cusignal.firwin(numtaps, f, window='nuttall')
271
+ array([ 3.56607041e-04, 9.99286786e-01, 3.56607041e-04])
272
+
273
+ High-pass ('stop' from 0 to f):
274
+
275
+ >>> cusignal.firwin(numtaps, f, pass_zero=False)
276
+ array([-0.00859313, 0.98281375, -0.00859313])
277
+
278
+ Band-pass:
279
+
280
+ >>> f1, f2 = 0.1, 0.2
281
+ >>> cusignal.firwin(numtaps, [f1, f2], pass_zero=False)
282
+ array([ 0.06301614, 0.88770441, 0.06301614])
283
+
284
+ Band-stop:
285
+
286
+ >>> cusignal.firwin(numtaps, [f1, f2])
287
+ array([-0.00801395, 1.0160279 , -0.00801395])
288
+
289
+ Multi-band (passbands are [0, f1], [f2, f3] and [f4, 1]):
290
+
291
+ >>> f3, f4 = 0.3, 0.4
292
+ >>> cusignal.firwin(numtaps, [f1, f2, f3, f4])
293
+ array([-0.01376344, 1.02752689, -0.01376344])
294
+
295
+ Multi-band (passbands are [f1, f2] and [f3,f4]):
296
+
297
+ >>> cusignal.firwin(numtaps, [f1, f2, f3, f4], pass_zero=False)
298
+ array([ 0.04890915, 0.91284326, 0.04890915])
299
+
300
+ """
301
+
302
+ nyq = 0.5 * fs
303
+
304
+ cutoff = cupy.atleast_1d(cutoff) / float(nyq)
305
+
306
+ # Check for invalid input.
307
+ if cutoff.ndim > 1:
308
+ raise ValueError(
309
+ "The cutoff argument must be at most " "one-dimensional.")
310
+ if cutoff.size == 0:
311
+ raise ValueError("At least one cutoff frequency must be given.")
312
+ if cutoff.min() <= 0 or cutoff.max() >= 1:
313
+ raise ValueError(
314
+ "Invalid cutoff frequency: frequencies must be "
315
+ "greater than 0 and less than nyq."
316
+ )
317
+ if cupy.any(cupy.diff(cutoff) <= 0):
318
+ raise ValueError(
319
+ "Invalid cutoff frequencies: the frequencies "
320
+ "must be strictly increasing."
321
+ )
322
+
323
+ if width is not None:
324
+ # A width was given. Find the beta parameter of the Kaiser window
325
+ # and set `window`. This overrides the value of `window` passed in.
326
+ atten = kaiser_atten(numtaps, float(width) / nyq)
327
+ beta = kaiser_beta(atten)
328
+ window = ("kaiser", beta)
329
+
330
+ if isinstance(pass_zero, str):
331
+ if pass_zero in ("bandstop", "lowpass"):
332
+ if pass_zero == "lowpass":
333
+ if cutoff.size != 1:
334
+ raise ValueError(
335
+ "cutoff must have one element if "
336
+ 'pass_zero=="lowpass", got %s' % (cutoff.shape,)
337
+ )
338
+ elif cutoff.size <= 1:
339
+ raise ValueError(
340
+ "cutoff must have at least two elements if "
341
+ 'pass_zero=="bandstop", got %s' % (cutoff.shape,)
342
+ )
343
+ pass_zero = True
344
+ elif pass_zero in ("bandpass", "highpass"):
345
+ if pass_zero == "highpass":
346
+ if cutoff.size != 1:
347
+ raise ValueError(
348
+ "cutoff must have one element if "
349
+ 'pass_zero=="highpass", got %s' % (cutoff.shape,)
350
+ )
351
+ elif cutoff.size <= 1:
352
+ raise ValueError(
353
+ "cutoff must have at least two elements if "
354
+ 'pass_zero=="bandpass", got %s' % (cutoff.shape,)
355
+ )
356
+ pass_zero = False
357
+ else:
358
+ raise ValueError(
359
+ 'pass_zero must be True, False, "bandpass", '
360
+ '"lowpass", "highpass", or "bandstop", got '
361
+ "{}".format(pass_zero)
362
+ )
363
+
364
+ pass_nyquist = bool(cutoff.size & 1) ^ pass_zero
365
+
366
+ if pass_nyquist and numtaps % 2 == 0:
367
+ raise ValueError(
368
+ "A filter with an even number of coefficients must "
369
+ "have zero response at the Nyquist rate."
370
+ )
371
+
372
+ # Insert 0 and/or 1 at the ends of cutoff so that the length of cutoff
373
+ # is even, and each pair in cutoff corresponds to passband.
374
+ cutoff = cupy.hstack(([0.0] * pass_zero, cutoff, [1.0] * pass_nyquist))
375
+
376
+ # `bands` is a 2D array; each row gives the left and right edges of
377
+ # a passband.
378
+ bands = cutoff.reshape(-1, 2)
379
+
380
+ win = get_window(window, numtaps, fftbins=False)
381
+ h, hc = _firwin_kernel(win, numtaps, bands, bands.shape[0], scale)
382
+ if scale:
383
+ s = cupy.sum(hc)
384
+ h /= s
385
+
386
+ # Build up the coefficients.
387
+ alpha = 0.5 * (numtaps - 1)
388
+ m = cupy.arange(0, numtaps) - alpha
389
+ h = 0
390
+ for left, right in bands:
391
+ h += right * cupy.sinc(right * m)
392
+ h -= left * cupy.sinc(left * m)
393
+
394
+ h *= win
395
+
396
+ # Now handle scaling if desired.
397
+ if scale:
398
+ # Get the first passband.
399
+ left, right = bands[0]
400
+ if left == 0:
401
+ scale_frequency = 0.0
402
+ elif right == 1:
403
+ scale_frequency = 1.0
404
+ else:
405
+ scale_frequency = 0.5 * (left + right)
406
+ c = cupy.cos(cupy.pi * m * scale_frequency)
407
+ s = cupy.sum(h * c)
408
+ h /= s
409
+
410
+ return h
411
+
412
+
413
+ def firwin2(
414
+ numtaps,
415
+ freq,
416
+ gain,
417
+ nfreqs=None,
418
+ window="hamming",
419
+ nyq=None,
420
+ antisymmetric=False,
421
+ fs=2.0,
422
+ ):
423
+ """
424
+ FIR filter design using the window method.
425
+
426
+ From the given frequencies `freq` and corresponding gains `gain`,
427
+ this function constructs an FIR filter with linear phase and
428
+ (approximately) the given frequency response.
429
+
430
+ Parameters
431
+ ----------
432
+ numtaps : int
433
+ The number of taps in the FIR filter. `numtaps` must be less than
434
+ `nfreqs`.
435
+ freq : array_like, 1-D
436
+ The frequency sampling points. Typically 0.0 to 1.0 with 1.0 being
437
+ Nyquist. The Nyquist frequency is half `fs`.
438
+ The values in `freq` must be nondecreasing. A value can be repeated
439
+ once to implement a discontinuity. The first value in `freq` must
440
+ be 0, and the last value must be ``fs/2``. Values 0 and ``fs/2`` must
441
+ not be repeated.
442
+ gain : array_like
443
+ The filter gains at the frequency sampling points. Certain
444
+ constraints to gain values, depending on the filter type, are applied,
445
+ see Notes for details.
446
+ nfreqs : int, optional
447
+ The size of the interpolation mesh used to construct the filter.
448
+ For most efficient behavior, this should be a power of 2 plus 1
449
+ (e.g, 129, 257, etc). The default is one more than the smallest
450
+ power of 2 that is not less than `numtaps`. `nfreqs` must be greater
451
+ than `numtaps`.
452
+ window : string or (string, float) or float, or None, optional
453
+ Window function to use. Default is "hamming". See
454
+ `scipy.signal.get_window` for the complete list of possible values.
455
+ If None, no window function is applied.
456
+ antisymmetric : bool, optional
457
+ Whether resulting impulse response is symmetric/antisymmetric.
458
+ See Notes for more details.
459
+ fs : float, optional
460
+ The sampling frequency of the signal. Each frequency in `cutoff`
461
+ must be between 0 and ``fs/2``. Default is 2.
462
+
463
+ Returns
464
+ -------
465
+ taps : ndarray
466
+ The filter coefficients of the FIR filter, as a 1-D array of length
467
+ `numtaps`.
468
+
469
+ See Also
470
+ --------
471
+ scipy.signal.firwin2
472
+ firls
473
+ firwin
474
+ minimum_phase
475
+ remez
476
+
477
+ Notes
478
+ -----
479
+ From the given set of frequencies and gains, the desired response is
480
+ constructed in the frequency domain. The inverse FFT is applied to the
481
+ desired response to create the associated convolution kernel, and the
482
+ first `numtaps` coefficients of this kernel, scaled by `window`, are
483
+ returned.
484
+ The FIR filter will have linear phase. The type of filter is determined by
485
+ the value of 'numtaps` and `antisymmetric` flag.
486
+ There are four possible combinations:
487
+
488
+ - odd `numtaps`, `antisymmetric` is False, type I filter is produced
489
+ - even `numtaps`, `antisymmetric` is False, type II filter is produced
490
+ - odd `numtaps`, `antisymmetric` is True, type III filter is produced
491
+ - even `numtaps`, `antisymmetric` is True, type IV filter is produced
492
+
493
+ Magnitude response of all but type I filters are subjects to following
494
+ constraints:
495
+
496
+ - type II -- zero at the Nyquist frequency
497
+ - type III -- zero at zero and Nyquist frequencies
498
+ - type IV -- zero at zero frequency
499
+ """
500
+ nyq = 0.5 * fs
501
+
502
+ if len(freq) != len(gain):
503
+ raise ValueError("freq and gain must be of same length.")
504
+
505
+ if nfreqs is not None and numtaps >= nfreqs:
506
+ raise ValueError(
507
+ (
508
+ "ntaps must be less than nfreqs, but firwin2 was "
509
+ "called with ntaps=%d and nfreqs=%s"
510
+ )
511
+ % (numtaps, nfreqs)
512
+ )
513
+
514
+ if freq[0] != 0 or freq[-1] != nyq:
515
+ raise ValueError("freq must start with 0 and end with fs/2.")
516
+ d = cupy.diff(freq)
517
+ if (d < 0).any():
518
+ raise ValueError("The values in freq must be nondecreasing.")
519
+ d2 = d[:-1] + d[1:]
520
+ if (d2 == 0).any():
521
+ raise ValueError("A value in freq must not occur more than twice.")
522
+ if freq[1] == 0:
523
+ raise ValueError("Value 0 must not be repeated in freq")
524
+ if freq[-2] == nyq:
525
+ raise ValueError("Value fs/2 must not be repeated in freq")
526
+
527
+ if antisymmetric:
528
+ if numtaps % 2 == 0:
529
+ ftype = 4
530
+ else:
531
+ ftype = 3
532
+ else:
533
+ if numtaps % 2 == 0:
534
+ ftype = 2
535
+ else:
536
+ ftype = 1
537
+
538
+ if ftype == 2 and gain[-1] != 0.0:
539
+ raise ValueError(
540
+ "A Type II filter must have zero gain at the " "Nyquist frequency."
541
+ )
542
+ elif ftype == 3 and (gain[0] != 0.0 or gain[-1] != 0.0):
543
+ raise ValueError(
544
+ "A Type III filter must have zero gain at zero "
545
+ "and Nyquist frequencies."
546
+ )
547
+ elif ftype == 4 and gain[0] != 0.0:
548
+ raise ValueError(
549
+ "A Type IV filter must have zero gain at zero " "frequency.")
550
+
551
+ if nfreqs is None:
552
+ nfreqs = 1 + 2 ** int(math.ceil(math.log(numtaps, 2)))
553
+
554
+ if (d == 0).any():
555
+ # Tweak any repeated values in freq so that interp works.
556
+ freq = cupy.array(freq, copy=True)
557
+ eps = cupy.finfo(float).eps * nyq
558
+ for k in range(len(freq) - 1):
559
+ if freq[k] == freq[k + 1]:
560
+ freq[k] = freq[k] - eps
561
+ freq[k + 1] = freq[k + 1] + eps
562
+ # Check if freq is strictly increasing after tweak
563
+ d = cupy.diff(freq)
564
+ if (d <= 0).any():
565
+ raise ValueError(
566
+ "freq cannot contain numbers that are too close "
567
+ "(within eps * (fs/2): "
568
+ "{}) to a repeated value".format(eps)
569
+ )
570
+
571
+ # Linearly interpolate the desired response on a uniform mesh `x`.
572
+ x = cupy.linspace(0.0, nyq, nfreqs)
573
+ fx = cupy.interp(x, freq, gain)
574
+
575
+ # Adjust the phases of the coefficients so that the first `ntaps` of the
576
+ # inverse FFT are the desired filter coefficients.
577
+ shift = cupy.exp(-(numtaps - 1) / 2.0 * 1.0j * math.pi * x / nyq)
578
+ if ftype > 2:
579
+ shift *= 1j
580
+
581
+ fx2 = fx * shift
582
+
583
+ # Use irfft to compute the inverse FFT.
584
+ out_full = cupy.fft.irfft(fx2)
585
+
586
+ if window is not None:
587
+ # Create the window to apply to the filter coefficients.
588
+ wind = get_window(window, numtaps, fftbins=False)
589
+ else:
590
+ wind = 1
591
+
592
+ # Keep only the first `numtaps` coefficients in `out`, and multiply by
593
+ # the window.
594
+ out = out_full[:numtaps] * wind
595
+
596
+ if ftype == 3:
597
+ out[out.size // 2] = 0.0
598
+
599
+ return out
600
+
601
+
602
+ # Scipy <= 1.12 has a deprecated `nyq` argument (nyq = fs/2).
603
+ # Remove it here, to be forward-looking.
604
+ def firls(numtaps, bands, desired, weight=None, fs=2):
605
+ """
606
+ FIR filter design using least-squares error minimization.
607
+
608
+ Calculate the filter coefficients for the linear-phase finite
609
+ impulse response (FIR) filter which has the best approximation
610
+ to the desired frequency response described by `bands` and
611
+ `desired` in the least squares sense (i.e., the integral of the
612
+ weighted mean-squared error within the specified bands is
613
+ minimized).
614
+
615
+ Parameters
616
+ ----------
617
+ numtaps : int
618
+ The number of taps in the FIR filter. `numtaps` must be odd.
619
+ bands : array_like
620
+ A monotonic nondecreasing sequence containing the band edges in
621
+ Hz. All elements must be non-negative and less than or equal to
622
+ the Nyquist frequency given by `fs`/2. The bands are specified as
623
+ frequency pairs, thus, if using a 1D array, its length must be
624
+ even, e.g., `cupy.array([0, 1, 2, 3, 4, 5])`. Alternatively, the
625
+ bands can be specified as an nx2 sized 2D array, where n is the
626
+ number of bands, e.g, `cupy.array([[0, 1], [2, 3], [4, 5]])`.
627
+ All elements of `bands` must be monotonically nondecreasing, have
628
+ width > 0, and must not overlap. (This is not checked by the routine).
629
+ desired : array_like
630
+ A sequence the same size as `bands` containing the desired gain
631
+ at the start and end point of each band.
632
+ All elements must be non-negative (this is not checked by the routine).
633
+ weight : array_like, optional
634
+ A relative weighting to give to each band region when solving
635
+ the least squares problem. `weight` has to be half the size of
636
+ `bands`.
637
+ All elements must be non-negative (this is not checked by the routine).
638
+ fs : float, optional
639
+ The sampling frequency of the signal. Each frequency in `bands`
640
+ must be between 0 and ``fs/2`` (inclusive). Default is 2.
641
+
642
+ Returns
643
+ -------
644
+ coeffs : ndarray
645
+ Coefficients of the optimal (in a least squares sense) FIR filter.
646
+
647
+ See Also
648
+ --------
649
+ firwin
650
+ firwin2
651
+ minimum_phase
652
+ remez
653
+ scipy.signal.firls
654
+ """
655
+ nyq = 0.5 * fs
656
+
657
+ numtaps = int(numtaps)
658
+ if numtaps % 2 == 0 or numtaps < 1:
659
+ raise ValueError("numtaps must be odd and >= 1")
660
+ M = (numtaps-1) // 2
661
+
662
+ # normalize bands 0->1 and make it 2 columns
663
+ nyq = float(nyq)
664
+ if nyq <= 0:
665
+ raise ValueError('nyq must be positive, got %s <= 0.' % nyq)
666
+ bands = cupy.asarray(bands).flatten() / nyq
667
+ if len(bands) % 2 != 0:
668
+ raise ValueError("bands must contain frequency pairs.")
669
+ if (bands < 0).any() or (bands > 1).any():
670
+ raise ValueError("bands must be between 0 and 1 relative to Nyquist")
671
+ bands.shape = (-1, 2)
672
+
673
+ # check remaining params
674
+ desired = cupy.asarray(desired).flatten()
675
+ if bands.size != desired.size:
676
+ raise ValueError("desired must have one entry per frequency, got %s "
677
+ "gains for %s frequencies."
678
+ % (desired.size, bands.size))
679
+ desired.shape = (-1, 2)
680
+ # if (cupy.diff(bands) <= 0).any() or (cupy.diff(bands[:, 0]) < 0).any():
681
+ # raise ValueError("bands must be monotonically nondecreasing and have"
682
+ # " width > 0.")
683
+ # if (bands[:-1, 1] > bands[1:, 0]).any():
684
+ # raise ValueError("bands must not overlap.")
685
+ # if (desired < 0).any():
686
+ # raise ValueError("desired must be non-negative.")
687
+ if weight is None:
688
+ weight = cupy.ones(len(desired))
689
+ weight = cupy.asarray(weight).flatten()
690
+ if len(weight) != len(desired):
691
+ raise ValueError("weight must be the same size as the number of "
692
+ "band pairs ({}).".format(len(bands)))
693
+ # if (weight < 0).any():
694
+ # raise ValueError("weight must be non-negative.")
695
+
696
+ # Set up the linear matrix equation to be solved, Qa = b
697
+
698
+ # We can express Q(k,n) = 0.5 Q1(k,n) + 0.5 Q2(k,n)
699
+ # where Q1(k,n)=q(k-n) and Q2(k,n)=q(k+n), i.e. a Toeplitz plus Hankel.
700
+
701
+ # We omit the factor of 0.5 above, instead adding it during coefficient
702
+ # calculation.
703
+
704
+ # We also omit the 1/π from both Q and b equations, as they cancel
705
+ # during solving.
706
+
707
+ # We have that:
708
+ # q(n) = 1/π ∫W(ω)cos(nω)dω (over 0->π)
709
+ # Using our normalization ω=πf and with a constant weight W over each
710
+ # interval f1->f2 we get:
711
+ # q(n) = W∫cos(πnf)df (0->1) = Wf sin(πnf)/πnf
712
+ # integrated over each f1->f2 pair (i.e., value at f2 - value at f1).
713
+ n = cupy.arange(numtaps)[:, cupy.newaxis, cupy.newaxis]
714
+ q = cupy.dot(cupy.diff(cupy.sinc(bands * n) *
715
+ bands, axis=2)[:, :, 0], weight)
716
+
717
+ # Now we assemble our sum of Toeplitz and Hankel
718
+ Q1 = toeplitz(q[:M+1])
719
+ Q2 = hankel(q[:M+1], q[M:])
720
+ Q = Q1 + Q2
721
+
722
+ # Now for b(n) we have that:
723
+ # b(n) = 1/π ∫ W(ω)D(ω)cos(nω)dω (over 0->π)
724
+ # Using our normalization ω=πf and with a constant weight W over each
725
+ # interval and a linear term for D(ω) we get (over each f1->f2 interval):
726
+ # b(n) = W ∫ (mf+c)cos(πnf)df
727
+ # = f(mf+c)sin(πnf)/πnf + mf**2 cos(nπf)/(πnf)**2
728
+ # integrated over each f1->f2 pair (i.e., value at f2 - value at f1).
729
+ n = n[:M + 1] # only need this many coefficients here
730
+ # Choose m and c such that we are at the start and end weights
731
+ m = (cupy.diff(desired, axis=1) / cupy.diff(bands, axis=1))
732
+ c = desired[:, [0]] - bands[:, [0]] * m
733
+ b = bands * (m*bands + c) * cupy.sinc(bands * n)
734
+ # Use L'Hospital's rule here for cos(nπf)/(πnf)**2 @ n=0
735
+ b[0] -= m * bands * bands / 2.
736
+ b[1:] += m * cupy.cos(n[1:] * cupy.pi * bands) / (cupy.pi * n[1:]) ** 2
737
+ b = cupy.diff(b, axis=2)[:, :, 0] @ weight
738
+
739
+ # Now we can solve the equation : XXX CuPy failure modes (?)
740
+ with cupyx.errstate(linalg="raise"):
741
+ try:
742
+ a = solve(Q, b)
743
+ except LinAlgError:
744
+ # in case Q is rank deficient
745
+ a = lstsq(Q, b, rcond=None)[0]
746
+
747
+ # XXX: scipy.signal does this:
748
+ # try: # try the fast way
749
+ # with warnings.catch_warnings(record=True) as w:
750
+ # warnings.simplefilter('always')
751
+ # a = solve(Q, b)
752
+ # for ww in w:
753
+ # if (ww.category == LinAlgWarning and
754
+ # str(ww.message).startswith('Ill-conditioned matrix')):
755
+ # raise LinAlgError(str(ww.message))
756
+ # except LinAlgError: # in case Q is rank deficient
757
+ # a = lstsq(Q, b)[0]
758
+
759
+ # make coefficients symmetric (linear phase)
760
+ coeffs = cupy.hstack((a[:0:-1], 2 * a[0], a[1:]))
761
+ return coeffs
762
+
763
+
764
+ def _dhtm(mag):
765
+ """Compute the modified 1-D discrete Hilbert transform
766
+
767
+ Parameters
768
+ ----------
769
+ mag : ndarray
770
+ The magnitude spectrum. Should be 1-D with an even length, and
771
+ preferably a fast length for FFT/IFFT.
772
+ """
773
+ # Adapted based on code by Niranjan Damera-Venkata,
774
+ # Brian L. Evans and Shawn R. McCaslin (see refs for `minimum_phase`)
775
+ sig = cupy.zeros(len(mag))
776
+ # Leave Nyquist and DC at 0, knowing np.abs(fftfreq(N)[midpt]) == 0.5
777
+ midpt = len(mag) // 2
778
+ sig[1:midpt] = 1
779
+ sig[midpt + 1:] = -1
780
+ # eventually if we want to support complex filters, we will need a
781
+ # cupy.abs() on the mag inside the log, and should remove the .real
782
+ recon = ifft(mag * cupy.exp(fft(sig * ifft(cupy.log(mag))))).real
783
+ return recon
784
+
785
+
786
+ def minimum_phase(h, method='homomorphic', n_fft=None):
787
+ """Convert a linear-phase FIR filter to minimum phase
788
+
789
+ Parameters
790
+ ----------
791
+ h : array
792
+ Linear-phase FIR filter coefficients.
793
+ method : {'hilbert', 'homomorphic'}
794
+ The method to use:
795
+
796
+ 'homomorphic' (default)
797
+ This method [4]_ [5]_ works best with filters with an
798
+ odd number of taps, and the resulting minimum phase filter
799
+ will have a magnitude response that approximates the square
800
+ root of the original filter's magnitude response.
801
+
802
+ 'hilbert'
803
+ This method [1]_ is designed to be used with equiripple
804
+ filters (e.g., from `remez`) with unity or zero gain
805
+ regions.
806
+
807
+ n_fft : int
808
+ The number of points to use for the FFT. Should be at least a
809
+ few times larger than the signal length (see Notes).
810
+
811
+ Returns
812
+ -------
813
+ h_minimum : array
814
+ The minimum-phase version of the filter, with length
815
+ ``(length(h) + 1) // 2``.
816
+
817
+ See Also
818
+ --------
819
+ scipy.signal.minimum_phase
820
+
821
+ Notes
822
+ -----
823
+ Both the Hilbert [1]_ or homomorphic [4]_ [5]_ methods require selection
824
+ of an FFT length to estimate the complex cepstrum of the filter.
825
+
826
+ In the case of the Hilbert method, the deviation from the ideal
827
+ spectrum ``epsilon`` is related to the number of stopband zeros
828
+ ``n_stop`` and FFT length ``n_fft`` as::
829
+
830
+ epsilon = 2. * n_stop / n_fft
831
+
832
+ For example, with 100 stopband zeros and a FFT length of 2048,
833
+ ``epsilon = 0.0976``. If we conservatively assume that the number of
834
+ stopband zeros is one less than the filter length, we can take the FFT
835
+ length to be the next power of 2 that satisfies ``epsilon=0.01`` as::
836
+
837
+ n_fft = 2 ** int(np.ceil(np.log2(2 * (len(h) - 1) / 0.01)))
838
+
839
+ This gives reasonable results for both the Hilbert and homomorphic
840
+ methods, and gives the value used when ``n_fft=None``.
841
+
842
+ Alternative implementations exist for creating minimum-phase filters,
843
+ including zero inversion [2]_ and spectral factorization [3]_ [4]_ [5]_.
844
+ For more information, see:
845
+
846
+ http://dspguru.com/dsp/howtos/how-to-design-minimum-phase-fir-filters
847
+
848
+ References
849
+ ----------
850
+ .. [1] N. Damera-Venkata and B. L. Evans, "Optimal design of real and
851
+ complex minimum phase digital FIR filters," Acoustics, Speech,
852
+ and Signal Processing, 1999. Proceedings., 1999 IEEE International
853
+ Conference on, Phoenix, AZ, 1999, pp. 1145-1148 vol.3.
854
+ DOI:10.1109/ICASSP.1999.756179
855
+ .. [2] X. Chen and T. W. Parks, "Design of optimal minimum phase FIR
856
+ filters by direct factorization," Signal Processing,
857
+ vol. 10, no. 4, pp. 369-383, Jun. 1986.
858
+ .. [3] T. Saramaki, "Finite Impulse Response Filter Design," in
859
+ Handbook for Digital Signal Processing, chapter 4,
860
+ New York: Wiley-Interscience, 1993.
861
+ .. [4] J. S. Lim, Advanced Topics in Signal Processing.
862
+ Englewood Cliffs, N.J.: Prentice Hall, 1988.
863
+ .. [5] A. V. Oppenheim, R. W. Schafer, and J. R. Buck,
864
+ "Discrete-Time Signal Processing," 2nd edition.
865
+ Upper Saddle River, N.J.: Prentice Hall, 1999.
866
+
867
+ """ # noqa
868
+ if cupy.iscomplexobj(h):
869
+ raise ValueError('Complex filters not supported')
870
+ if h.ndim != 1 or h.size <= 2:
871
+ raise ValueError('h must be 1-D and at least 2 samples long')
872
+ n_half = len(h) // 2
873
+ if not cupy.allclose(h[-n_half:][::-1], h[:n_half]):
874
+ import warnings
875
+ warnings.warn('h does not appear to by symmetric, conversion may '
876
+ 'fail', RuntimeWarning)
877
+ if not isinstance(method, str) or method not in \
878
+ ('homomorphic', 'hilbert',):
879
+ raise ValueError('method must be "homomorphic" or "hilbert", got %r'
880
+ % (method,))
881
+ if n_fft is None:
882
+ n_fft = 2 ** int(cupy.ceil(cupy.log2(2 * (len(h) - 1) / 0.01)))
883
+ n_fft = int(n_fft)
884
+ if n_fft < len(h):
885
+ raise ValueError('n_fft must be at least len(h)==%s' % len(h))
886
+ if method == 'hilbert':
887
+ w = cupy.arange(n_fft) * (2 * cupy.pi / n_fft * n_half)
888
+ H = cupy.real(fft(h, n_fft) * cupy.exp(1j * w))
889
+ dp = max(H) - 1
890
+ ds = 0 - min(H)
891
+ S = 4. / (cupy.sqrt(1 + dp + ds) + cupy.sqrt(1 - dp + ds)) ** 2
892
+ H += ds
893
+ H *= S
894
+ H = cupy.sqrt(H, out=H)
895
+ H += 1e-10 # ensure that the log does not explode
896
+ h_minimum = _dhtm(H)
897
+ else: # method == 'homomorphic'
898
+ # zero-pad; calculate the DFT
899
+ h_temp = cupy.abs(fft(h, n_fft))
900
+ # take 0.25*log(|H|**2) = 0.5*log(|H|)
901
+ h_temp += 1e-7 * h_temp[h_temp > 0].min() # don't let log blow up
902
+ cupy.log(h_temp, out=h_temp)
903
+ h_temp *= 0.5
904
+ # IDFT
905
+ h_temp = ifft(h_temp).real
906
+ # multiply pointwise by the homomorphic filter
907
+ # lmin[n] = 2u[n] - d[n]
908
+ win = cupy.zeros(n_fft)
909
+ win[0] = 1
910
+ stop = (len(h) + 1) // 2
911
+ win[1:stop] = 2
912
+ if len(h) % 2:
913
+ win[stop] = 1
914
+ h_temp *= win
915
+ h_temp = ifft(cupy.exp(fft(h_temp)))
916
+ h_minimum = h_temp.real
917
+ n_out = n_half + len(h) % 2
918
+ return h_minimum[:n_out]
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_iir_filter_conversions.py ADDED
@@ -0,0 +1,2364 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """ IIR filter conversion utilities.
2
+
3
+ Split off _filter_design.py
4
+ """
5
+ import warnings
6
+ import math
7
+ from math import pi, prod
8
+
9
+ import cupy
10
+ from cupyx.scipy.special import binom as comb
11
+ import cupyx.scipy.special as special
12
+ from cupyx.scipy.signal import _optimize
13
+
14
+ from cupyx.scipy.signal._polyutils import roots, poly
15
+ from cupyx.scipy.signal._lti_conversion import abcd_normalize
16
+
17
+
18
+ class BadCoefficients(UserWarning):
19
+ """Warning about badly conditioned filter coefficients"""
20
+ pass
21
+
22
+
23
+ def _trim_zeros(filt, trim='fb'):
24
+ # https://github.com/numpy/numpy/blob/v1.24.0/numpy/lib/function_base.py#L1800-L1850
25
+
26
+ first = 0
27
+ if 'f' in trim:
28
+ for i in filt:
29
+ if i != 0.:
30
+ break
31
+ else:
32
+ first = first + 1
33
+
34
+ last = len(filt)
35
+ if 'b' in trim:
36
+ for i in filt[::-1]:
37
+ if i != 0.:
38
+ break
39
+ else:
40
+ last = last - 1
41
+ return filt[first:last]
42
+
43
+
44
+ def _align_nums(nums):
45
+ """Aligns the shapes of multiple numerators.
46
+
47
+ Given an array of numerator coefficient arrays [[a_1, a_2,...,
48
+ a_n],..., [b_1, b_2,..., b_m]], this function pads shorter numerator
49
+ arrays with zero's so that all numerators have the same length. Such
50
+ alignment is necessary for functions like 'tf2ss', which needs the
51
+ alignment when dealing with SIMO transfer functions.
52
+
53
+ Parameters
54
+ ----------
55
+ nums: array_like
56
+ Numerator or list of numerators. Not necessarily with same length.
57
+
58
+ Returns
59
+ -------
60
+ nums: array
61
+ The numerator. If `nums` input was a list of numerators then a 2-D
62
+ array with padded zeros for shorter numerators is returned. Otherwise
63
+ returns ``np.asarray(nums)``.
64
+ """
65
+ try:
66
+ # The statement can throw a ValueError if one
67
+ # of the numerators is a single digit and another
68
+ # is array-like e.g. if nums = [5, [1, 2, 3]]
69
+ nums = cupy.asarray(nums)
70
+ return nums
71
+
72
+ except ValueError:
73
+ nums = [cupy.atleast_1d(num) for num in nums]
74
+ max_width = max(num.size for num in nums)
75
+
76
+ # pre-allocate
77
+ aligned_nums = cupy.zeros((len(nums), max_width))
78
+
79
+ # Create numerators with padded zeros
80
+ for index, num in enumerate(nums):
81
+ aligned_nums[index, -num.size:] = num
82
+
83
+ return aligned_nums
84
+
85
+
86
+ def _polycoeffs_from_zeros(zeros, tol=10):
87
+ # a clone of numpy.poly, simplified
88
+ dtyp = (cupy.complex128
89
+ if cupy.issubdtype(zeros.dtype, cupy.complexfloating)
90
+ else cupy.float64)
91
+ a = cupy.ones(1, dtype=dtyp)
92
+ for z in zeros:
93
+ a = cupy.convolve(a, cupy.r_[1, -z], mode='full')
94
+
95
+ # Use real output if possible.
96
+ if dtyp == cupy.complex128:
97
+ mask = cupy.abs(a.imag) < tol * cupy.finfo(a.dtype).eps
98
+ a.imag[mask] = 0.0
99
+ if mask.shape[0] == a.shape[0]:
100
+ # all imag parts were fp noise
101
+ a = a.real.copy()
102
+ else:
103
+ # if all cmplx roots are complex conj, the coefficients are real
104
+ pos_roots = z[z.imag > 0]
105
+ neg_roots = z[z.imag < 0]
106
+ if pos_roots.shape[0] == neg_roots.shape[0]:
107
+ neg_roots = neg_roots.copy()
108
+ neg_roots.sort()
109
+ pos_roots = pos_roots.copy()
110
+ pos_roots.sort()
111
+ if (neg_roots == pos_roots.conj()).all():
112
+ a = a.real.copy()
113
+ return a
114
+
115
+
116
+ def _nearest_real_complex_idx(fro, to, which):
117
+ """Get the next closest real or complex element based on distance"""
118
+ assert which in ('real', 'complex', 'any')
119
+ order = cupy.argsort(cupy.abs(fro - to))
120
+ if which == 'any':
121
+ return order[0]
122
+ else:
123
+ mask = cupy.isreal(fro[order])
124
+ if which == 'complex':
125
+ mask = ~mask
126
+ return order[cupy.nonzero(mask)[0][0]]
127
+
128
+
129
+ def _single_zpksos(z, p, k):
130
+ """Create one second-order section from up to two zeros and poles"""
131
+ sos = cupy.zeros(6)
132
+ b, a = zpk2tf(cupy.asarray(z), cupy.asarray(p), k)
133
+ sos[3-len(b):3] = b
134
+ sos[6-len(a):6] = a
135
+ return sos
136
+
137
+
138
+ def zpk2sos(z, p, k, pairing=None, *, analog=False):
139
+ """Return second-order sections from zeros, poles, and gain of a system
140
+
141
+ Parameters
142
+ ----------
143
+ z : array_like
144
+ Zeros of the transfer function.
145
+ p : array_like
146
+ Poles of the transfer function.
147
+ k : float
148
+ System gain.
149
+ pairing : {None, 'nearest', 'keep_odd', 'minimal'}, optional
150
+ The method to use to combine pairs of poles and zeros into sections.
151
+ If analog is False and pairing is None, pairing is set to 'nearest';
152
+ if analog is True, pairing must be 'minimal', and is set to that if
153
+ it is None.
154
+ analog : bool, optional
155
+ If True, system is analog, otherwise discrete.
156
+
157
+ Returns
158
+ -------
159
+ sos : ndarray
160
+ Array of second-order filter coefficients, with shape
161
+ ``(n_sections, 6)``. See `sosfilt` for the SOS filter format
162
+ specification.
163
+
164
+ See Also
165
+ --------
166
+ sosfilt
167
+ scipy.signal.zpk2sos
168
+
169
+ """
170
+ if pairing is None:
171
+ pairing = 'minimal' if analog else 'nearest'
172
+
173
+ valid_pairings = ['nearest', 'keep_odd', 'minimal']
174
+ if pairing not in valid_pairings:
175
+ raise ValueError('pairing must be one of %s, not %s'
176
+ % (valid_pairings, pairing))
177
+
178
+ if analog and pairing != 'minimal':
179
+ raise ValueError('for analog zpk2sos conversion, '
180
+ 'pairing must be "minimal"')
181
+
182
+ if len(z) == len(p) == 0:
183
+ if not analog:
184
+ return cupy.array([[k, 0., 0., 1., 0., 0.]])
185
+ else:
186
+ return cupy.array([[0., 0., k, 0., 0., 1.]])
187
+
188
+ if pairing != 'minimal':
189
+ # ensure we have the same number of poles and zeros, and make copies
190
+ p = cupy.concatenate((p, cupy.zeros(max(len(z) - len(p), 0))))
191
+ z = cupy.concatenate((z, cupy.zeros(max(len(p) - len(z), 0))))
192
+ n_sections = (max(len(p), len(z)) + 1) // 2
193
+
194
+ if len(p) % 2 == 1 and pairing == 'nearest':
195
+ p = cupy.concatenate((p, cupy.zeros(1)))
196
+ z = cupy.concatenate((z, cupy.zeros(1)))
197
+ assert len(p) == len(z)
198
+ else:
199
+ if len(p) < len(z):
200
+ raise ValueError('for analog zpk2sos conversion, '
201
+ 'must have len(p)>=len(z)')
202
+
203
+ n_sections = (len(p) + 1) // 2
204
+
205
+ # Ensure we have complex conjugate pairs
206
+ # (note that _cplxreal only gives us one element of each complex pair):
207
+ z = cupy.concatenate(_cplxreal(z))
208
+ p = cupy.concatenate(_cplxreal(p))
209
+ if not cupy.isreal(k):
210
+ raise ValueError('k must be real')
211
+ k = k.real
212
+
213
+ if not analog:
214
+ # digital: "worst" is the closest to the unit circle
215
+ def idx_worst(p):
216
+ return cupy.argmin(cupy.abs(1 - cupy.abs(p)))
217
+ else:
218
+ # analog: "worst" is the closest to the imaginary axis
219
+ def idx_worst(p):
220
+ return cupy.argmin(cupy.abs(cupy.real(p)))
221
+
222
+ sos = cupy.zeros((n_sections, 6))
223
+
224
+ # Construct the system, reversing order so the "worst" are last
225
+ for si in range(n_sections-1, -1, -1):
226
+ # Select the next "worst" pole
227
+ p1_idx = idx_worst(p)
228
+ p1 = p[p1_idx]
229
+ p = cupy.delete(p, p1_idx)
230
+
231
+ # Pair that pole with a zero
232
+
233
+ if cupy.isreal(p1) and cupy.isreal(p).sum() == 0:
234
+ # Special case (1): last remaining real pole
235
+ if pairing != 'minimal':
236
+ z1_idx = _nearest_real_complex_idx(z, p1, 'real')
237
+ z1 = z[z1_idx]
238
+ z = cupy.delete(z, z1_idx)
239
+ sos[si] = _single_zpksos(cupy.r_[z1, 0], cupy.r_[p1, 0], 1)
240
+ elif len(z) > 0:
241
+ z1_idx = _nearest_real_complex_idx(z, p1, 'real')
242
+ z1 = z[z1_idx]
243
+ z = cupy.delete(z, z1_idx)
244
+ sos[si] = _single_zpksos([z1], [p1], 1)
245
+ else:
246
+ sos[si] = _single_zpksos([], [p1], 1)
247
+
248
+ elif (len(p) + 1 == len(z)
249
+ and not cupy.isreal(p1)
250
+ and cupy.isreal(p).sum() == 1
251
+ and cupy.isreal(z).sum() == 1):
252
+
253
+ # Special case (2): there's one real pole and one real zero
254
+ # left, and an equal number of poles and zeros to pair up.
255
+ # We *must* pair with a complex zero
256
+
257
+ z1_idx = _nearest_real_complex_idx(z, p1, 'complex')
258
+ z1 = z[z1_idx]
259
+ z = cupy.delete(z, z1_idx)
260
+ sos[si] = _single_zpksos(
261
+ cupy.r_[z1, z1.conj()], cupy.r_[p1, p1.conj()], 1)
262
+
263
+ else:
264
+ if cupy.isreal(p1):
265
+ prealidx = cupy.flatnonzero(cupy.isreal(p))
266
+ p2_idx = prealidx[idx_worst(p[prealidx])]
267
+ p2 = p[p2_idx]
268
+ p = cupy.delete(p, p2_idx)
269
+ else:
270
+ p2 = p1.conj()
271
+
272
+ # find closest zero
273
+ if len(z) > 0:
274
+ z1_idx = _nearest_real_complex_idx(z, p1, 'any')
275
+ z1 = z[z1_idx]
276
+ z = cupy.delete(z, z1_idx)
277
+
278
+ if not cupy.isreal(z1):
279
+ sos[si] = _single_zpksos(
280
+ cupy.r_[z1, z1.conj()], cupy.r_[p1, p2], 1)
281
+ else:
282
+ if len(z) > 0:
283
+ z2_idx = _nearest_real_complex_idx(z, p1, 'real')
284
+ z2 = z[z2_idx]
285
+ assert cupy.isreal(z2)
286
+ z = cupy.delete(z, z2_idx)
287
+ sos[si] = _single_zpksos(cupy.r_[z1, z2], [p1, p2], 1)
288
+ else:
289
+ sos[si] = _single_zpksos([z1], [p1, p2], 1)
290
+ else:
291
+ # no more zeros
292
+ sos[si] = _single_zpksos([], [p1, p2], 1)
293
+
294
+ assert len(p) == len(z) == 0 # we've consumed all poles and zeros
295
+ del p, z
296
+
297
+ # put gain in first sos
298
+ sos[0][:3] *= k
299
+ return sos
300
+
301
+
302
+ def _cplxreal(z, tol=None):
303
+ """
304
+ Split into complex and real parts, combining conjugate pairs.
305
+
306
+ The 1-D input vector `z` is split up into its complex (zc) and real (zr)
307
+ elements. Every complex element must be part of a complex-conjugate pair,
308
+ which are combined into a single number (with positive imaginary part) in
309
+ the output. Two complex numbers are considered a conjugate pair if their
310
+ real and imaginary parts differ in magnitude by less than ``tol * abs(z)``.
311
+
312
+ Parameters
313
+ ----------
314
+ z : array_like
315
+ Vector of complex numbers to be sorted and split
316
+ tol : float, optional
317
+ Relative tolerance for testing realness and conjugate equality.
318
+ Default is ``100 * spacing(1)`` of `z`'s data type (i.e., 2e-14 for
319
+ float64)
320
+
321
+ Returns
322
+ -------
323
+ zc : ndarray
324
+ Complex elements of `z`, with each pair represented by a single value
325
+ having positive imaginary part, sorted first by real part, and then
326
+ by magnitude of imaginary part. The pairs are averaged when combined
327
+ to reduce error.
328
+ zr : ndarray
329
+ Real elements of `z` (those having imaginary part less than
330
+ `tol` times their magnitude), sorted by value.
331
+
332
+ Raises
333
+ ------
334
+ ValueError
335
+ If there are any complex numbers in `z` for which a conjugate
336
+ cannot be found.
337
+
338
+ See Also
339
+ --------
340
+ scipy.signal.cmplxreal
341
+
342
+ Examples
343
+ --------
344
+ >>> a = [4, 3, 1, 2-2j, 2+2j, 2-1j, 2+1j, 2-1j, 2+1j, 1+1j, 1-1j]
345
+ >>> zc, zr = _cplxreal(a)
346
+ >>> print(zc)
347
+ [ 1.+1.j 2.+1.j 2.+1.j 2.+2.j]
348
+ >>> print(zr)
349
+ [ 1. 3. 4.]
350
+ """
351
+
352
+ z = cupy.atleast_1d(z)
353
+ if z.size == 0:
354
+ return z, z
355
+ elif z.ndim != 1:
356
+ raise ValueError('_cplxreal only accepts 1-D input')
357
+
358
+ if tol is None:
359
+ # Get tolerance from dtype of input
360
+ tol = 100 * cupy.finfo((1.0 * z).dtype).eps
361
+
362
+ # Sort by real part, magnitude of imaginary part (speed up further sorting)
363
+ z = z[cupy.lexsort(cupy.array([abs(z.imag), z.real]))]
364
+
365
+ # Split reals from conjugate pairs
366
+ real_indices = abs(z.imag) <= tol * abs(z)
367
+ zr = z[real_indices].real
368
+
369
+ if len(zr) == len(z):
370
+ # Input is entirely real
371
+ return cupy.array([]), zr
372
+
373
+ # Split positive and negative halves of conjugates
374
+ z = z[~real_indices]
375
+ zp = z[z.imag > 0]
376
+ zn = z[z.imag < 0]
377
+
378
+ if len(zp) != len(zn):
379
+ raise ValueError('Array contains complex value with no matching '
380
+ 'conjugate.')
381
+
382
+ # Find runs of (approximately) the same real part
383
+ same_real = cupy.diff(zp.real) <= tol * abs(zp[:-1])
384
+ diffs = cupy.diff(cupy.r_[0, same_real, 0])
385
+ run_starts = cupy.nonzero(diffs > 0)[0]
386
+ run_stops = cupy.nonzero(diffs < 0)[0]
387
+
388
+ # Sort each run by their imaginary parts
389
+ for i in range(len(run_starts)):
390
+ start = run_starts[i]
391
+ stop = run_stops[i] + 1
392
+ for chunk in (zp[start:stop], zn[start:stop]):
393
+ chunk[...] = chunk[cupy.lexsort(cupy.array([abs(chunk.imag)]))]
394
+
395
+ # Check that negatives match positives
396
+ if any(abs(zp - zn.conj()) > tol * abs(zn)):
397
+ raise ValueError('Array contains complex value with no matching '
398
+ 'conjugate.')
399
+
400
+ # Average out numerical inaccuracy in real vs imag parts of pairs
401
+ zc = (zp + zn.conj()) / 2
402
+
403
+ return zc, zr
404
+
405
+
406
+ def normalize(b, a):
407
+ """Normalize numerator/denominator of a continuous-time transfer function.
408
+
409
+ If values of `b` are too close to 0, they are removed. In that case, a
410
+ BadCoefficients warning is emitted.
411
+
412
+ Parameters
413
+ ----------
414
+ b: array_like
415
+ Numerator of the transfer function. Can be a 2-D array to normalize
416
+ multiple transfer functions.
417
+ a: array_like
418
+ Denominator of the transfer function. At most 1-D.
419
+
420
+ Returns
421
+ -------
422
+ num: array
423
+ The numerator of the normalized transfer function. At least a 1-D
424
+ array. A 2-D array if the input `num` is a 2-D array.
425
+ den: 1-D array
426
+ The denominator of the normalized transfer function.
427
+
428
+ Notes
429
+ -----
430
+ Coefficients for both the numerator and denominator should be specified in
431
+ descending exponent order (e.g., ``s^2 + 3s + 5`` would be represented as
432
+ ``[1, 3, 5]``).
433
+
434
+ See Also
435
+ --------
436
+ scipy.signal.normalize
437
+
438
+ """
439
+ num, den = b, a
440
+
441
+ den = cupy.atleast_1d(den)
442
+ num = cupy.atleast_2d(_align_nums(num))
443
+
444
+ if den.ndim != 1:
445
+ raise ValueError("Denominator polynomial must be rank-1 array.")
446
+ if num.ndim > 2:
447
+ raise ValueError("Numerator polynomial must be rank-1 or"
448
+ " rank-2 array.")
449
+ if cupy.all(den == 0):
450
+ raise ValueError("Denominator must have at least on nonzero element.")
451
+
452
+ # Trim leading zeros in denominator, leave at least one.
453
+ den = _trim_zeros(den, 'f')
454
+
455
+ # Normalize transfer function
456
+ num, den = num / den[0], den / den[0]
457
+
458
+ # Count numerator columns that are all zero
459
+ leading_zeros = 0
460
+ for col in num.T:
461
+ if cupy.allclose(col, 0, atol=1e-14):
462
+ leading_zeros += 1
463
+ else:
464
+ break
465
+
466
+ # Trim leading zeros of numerator
467
+ if leading_zeros > 0:
468
+ warnings.warn("Badly conditioned filter coefficients (numerator): the "
469
+ "results may be meaningless", BadCoefficients)
470
+ # Make sure at least one column remains
471
+ if leading_zeros == num.shape[1]:
472
+ leading_zeros -= 1
473
+ num = num[:, leading_zeros:]
474
+
475
+ # Squeeze first dimension if singular
476
+ if num.shape[0] == 1:
477
+ num = num[0, :]
478
+
479
+ return num, den
480
+
481
+
482
+ def _relative_degree(z, p):
483
+ """
484
+ Return relative degree of transfer function from zeros and poles
485
+ """
486
+ degree = len(p) - len(z)
487
+ if degree < 0:
488
+ raise ValueError("Improper transfer function. "
489
+ "Must have at least as many poles as zeros.")
490
+ else:
491
+ return degree
492
+
493
+
494
+ def bilinear_zpk(z, p, k, fs):
495
+ r"""
496
+ Return a digital IIR filter from an analog one using a bilinear transform.
497
+
498
+ Transform a set of poles and zeros from the analog s-plane to the digital
499
+ z-plane using Tustin's method, which substitutes ``2*fs*(z-1) / (z+1)`` for
500
+ ``s``, maintaining the shape of the frequency response.
501
+
502
+ Parameters
503
+ ----------
504
+ z : array_like
505
+ Zeros of the analog filter transfer function.
506
+ p : array_like
507
+ Poles of the analog filter transfer function.
508
+ k : float
509
+ System gain of the analog filter transfer function.
510
+ fs : float
511
+ Sample rate, as ordinary frequency (e.g., hertz). No prewarping is
512
+ done in this function.
513
+
514
+ Returns
515
+ -------
516
+ z : ndarray
517
+ Zeros of the transformed digital filter transfer function.
518
+ p : ndarray
519
+ Poles of the transformed digital filter transfer function.
520
+ k : float
521
+ System gain of the transformed digital filter.
522
+
523
+ See Also
524
+ --------
525
+ lp2lp_zpk, lp2hp_zpk, lp2bp_zpk, lp2bs_zpk
526
+ bilinear
527
+ scipy.signal.bilinear_zpk
528
+
529
+ """
530
+ z = cupy.atleast_1d(z)
531
+ p = cupy.atleast_1d(p)
532
+
533
+ degree = _relative_degree(z, p)
534
+
535
+ fs2 = 2.0 * fs
536
+
537
+ # Bilinear transform the poles and zeros
538
+ z_z = (fs2 + z) / (fs2 - z)
539
+ p_z = (fs2 + p) / (fs2 - p)
540
+
541
+ # Any zeros that were at infinity get moved to the Nyquist frequency
542
+ z_z = cupy.append(z_z, -cupy.ones(degree))
543
+
544
+ # Compensate for gain change
545
+ k_z = k * (cupy.prod(fs2 - z) / cupy.prod(fs2 - p)).real
546
+
547
+ return z_z, p_z, k_z
548
+
549
+
550
+ def lp2lp_zpk(z, p, k, wo=1.0):
551
+ r"""
552
+ Transform a lowpass filter prototype to a different frequency.
553
+
554
+ Return an analog low-pass filter with cutoff frequency `wo`
555
+ from an analog low-pass filter prototype with unity cutoff frequency,
556
+ using zeros, poles, and gain ('zpk') representation.
557
+
558
+ Parameters
559
+ ----------
560
+ z : array_like
561
+ Zeros of the analog filter transfer function.
562
+ p : array_like
563
+ Poles of the analog filter transfer function.
564
+ k : float
565
+ System gain of the analog filter transfer function.
566
+ wo : float
567
+ Desired cutoff, as angular frequency (e.g., rad/s).
568
+ Defaults to no change.
569
+
570
+ Returns
571
+ -------
572
+ z : ndarray
573
+ Zeros of the transformed low-pass filter transfer function.
574
+ p : ndarray
575
+ Poles of the transformed low-pass filter transfer function.
576
+ k : float
577
+ System gain of the transformed low-pass filter.
578
+
579
+ See Also
580
+ --------
581
+ lp2hp_zpk, lp2bp_zpk, lp2bs_zpk, bilinear
582
+ lp2lp
583
+ scipy.signal.lp2lp_zpk
584
+
585
+ """
586
+ z = cupy.atleast_1d(z)
587
+ p = cupy.atleast_1d(p)
588
+ wo = float(wo) # Avoid int wraparound
589
+
590
+ degree = _relative_degree(z, p)
591
+
592
+ # Scale all points radially from origin to shift cutoff frequency
593
+ z_lp = wo * z
594
+ p_lp = wo * p
595
+
596
+ # Each shifted pole decreases gain by wo, each shifted zero increases it.
597
+ # Cancel out the net change to keep overall gain the same
598
+ k_lp = k * wo**degree
599
+
600
+ return z_lp, p_lp, k_lp
601
+
602
+
603
+ def lp2hp_zpk(z, p, k, wo=1.0):
604
+ r"""
605
+ Transform a lowpass filter prototype to a highpass filter.
606
+
607
+ Return an analog high-pass filter with cutoff frequency `wo`
608
+ from an analog low-pass filter prototype with unity cutoff frequency,
609
+ using zeros, poles, and gain ('zpk') representation.
610
+
611
+ Parameters
612
+ ----------
613
+ z : array_like
614
+ Zeros of the analog filter transfer function.
615
+ p : array_like
616
+ Poles of the analog filter transfer function.
617
+ k : float
618
+ System gain of the analog filter transfer function.
619
+ wo : float
620
+ Desired cutoff, as angular frequency (e.g., rad/s).
621
+ Defaults to no change.
622
+
623
+ Returns
624
+ -------
625
+ z : ndarray
626
+ Zeros of the transformed high-pass filter transfer function.
627
+ p : ndarray
628
+ Poles of the transformed high-pass filter transfer function.
629
+ k : float
630
+ System gain of the transformed high-pass filter.
631
+
632
+ See Also
633
+ --------
634
+ lp2lp_zpk, lp2bp_zpk, lp2bs_zpk, bilinear
635
+ lp2hp
636
+ scipy.signal.lp2hp_zpk
637
+
638
+ Notes
639
+ -----
640
+ This is derived from the s-plane substitution
641
+
642
+ .. math:: s \rightarrow \frac{\omega_0}{s}
643
+
644
+ This maintains symmetry of the lowpass and highpass responses on a
645
+ logarithmic scale.
646
+
647
+ """
648
+ z = cupy.atleast_1d(z)
649
+ p = cupy.atleast_1d(p)
650
+ wo = float(wo)
651
+
652
+ degree = _relative_degree(z, p)
653
+
654
+ # Invert positions radially about unit circle to convert LPF to HPF
655
+ # Scale all points radially from origin to shift cutoff frequency
656
+ z_hp = wo / z
657
+ p_hp = wo / p
658
+
659
+ # If lowpass had zeros at infinity, inverting moves them to origin.
660
+ z_hp = cupy.append(z_hp, cupy.zeros(degree))
661
+
662
+ # Cancel out gain change caused by inversion
663
+ k_hp = k * cupy.real(cupy.prod(-z) / cupy.prod(-p))
664
+
665
+ return z_hp, p_hp, k_hp
666
+
667
+
668
+ def lp2bp_zpk(z, p, k, wo=1.0, bw=1.0):
669
+ r"""
670
+ Transform a lowpass filter prototype to a bandpass filter.
671
+
672
+ Return an analog band-pass filter with center frequency `wo` and
673
+ bandwidth `bw` from an analog low-pass filter prototype with unity
674
+ cutoff frequency, using zeros, poles, and gain ('zpk') representation.
675
+
676
+ Parameters
677
+ ----------
678
+ z : array_like
679
+ Zeros of the analog filter transfer function.
680
+ p : array_like
681
+ Poles of the analog filter transfer function.
682
+ k : float
683
+ System gain of the analog filter transfer function.
684
+ wo : float
685
+ Desired passband center, as angular frequency (e.g., rad/s).
686
+ Defaults to no change.
687
+ bw : float
688
+ Desired passband width, as angular frequency (e.g., rad/s).
689
+ Defaults to 1.
690
+
691
+ Returns
692
+ -------
693
+ z : ndarray
694
+ Zeros of the transformed band-pass filter transfer function.
695
+ p : ndarray
696
+ Poles of the transformed band-pass filter transfer function.
697
+ k : float
698
+ System gain of the transformed band-pass filter.
699
+
700
+ See Also
701
+ --------
702
+ lp2lp_zpk, lp2hp_zpk, lp2bs_zpk, bilinear
703
+ lp2bp
704
+ scipy.signal.lp2bp_zpk
705
+
706
+ Notes
707
+ -----
708
+ This is derived from the s-plane substitution
709
+
710
+ .. math:: s \rightarrow \frac{s^2 + {\omega_0}^2}{s \cdot \mathrm{BW}}
711
+
712
+ This is the "wideband" transformation, producing a passband with
713
+ geometric (log frequency) symmetry about `wo`.
714
+
715
+ """
716
+ z = cupy.atleast_1d(z)
717
+ p = cupy.atleast_1d(p)
718
+ wo = float(wo)
719
+ bw = float(bw)
720
+
721
+ degree = _relative_degree(z, p)
722
+
723
+ # Scale poles and zeros to desired bandwidth
724
+ z_lp = z * bw/2
725
+ p_lp = p * bw/2
726
+
727
+ # Square root needs to produce complex result, not NaN
728
+ z_lp = z_lp.astype(complex)
729
+ p_lp = p_lp.astype(complex)
730
+
731
+ # Duplicate poles and zeros and shift from baseband to +wo and -wo
732
+ z_bp = cupy.concatenate((z_lp + cupy.sqrt(z_lp**2 - wo**2),
733
+ z_lp - cupy.sqrt(z_lp**2 - wo**2)))
734
+ p_bp = cupy.concatenate((p_lp + cupy.sqrt(p_lp**2 - wo**2),
735
+ p_lp - cupy.sqrt(p_lp**2 - wo**2)))
736
+
737
+ # Move degree zeros to origin, leaving degree zeros at infinity for BPF
738
+ z_bp = cupy.append(z_bp, cupy.zeros(degree))
739
+
740
+ # Cancel out gain change from frequency scaling
741
+ k_bp = k * bw**degree
742
+
743
+ return z_bp, p_bp, k_bp
744
+
745
+
746
+ def lp2bs_zpk(z, p, k, wo=1.0, bw=1.0):
747
+ r"""
748
+ Transform a lowpass filter prototype to a bandstop filter.
749
+
750
+ Return an analog band-stop filter with center frequency `wo` and
751
+ stopband width `bw` from an analog low-pass filter prototype with unity
752
+ cutoff frequency, using zeros, poles, and gain ('zpk') representation.
753
+
754
+ Parameters
755
+ ----------
756
+ z : array_like
757
+ Zeros of the analog filter transfer function.
758
+ p : array_like
759
+ Poles of the analog filter transfer function.
760
+ k : float
761
+ System gain of the analog filter transfer function.
762
+ wo : float
763
+ Desired stopband center, as angular frequency (e.g., rad/s).
764
+ Defaults to no change.
765
+ bw : float
766
+ Desired stopband width, as angular frequency (e.g., rad/s).
767
+ Defaults to 1.
768
+
769
+ Returns
770
+ -------
771
+ z : ndarray
772
+ Zeros of the transformed band-stop filter transfer function.
773
+ p : ndarray
774
+ Poles of the transformed band-stop filter transfer function.
775
+ k : float
776
+ System gain of the transformed band-stop filter.
777
+
778
+ See Also
779
+ --------
780
+ lp2lp_zpk, lp2hp_zpk, lp2bp_zpk, bilinear
781
+ lp2bs
782
+ scipy.signal.lp2bs_zpk
783
+
784
+ Notes
785
+ -----
786
+ This is derived from the s-plane substitution
787
+
788
+ .. math:: s \rightarrow \frac{s \cdot \mathrm{BW}}{s^2 + {\omega_0}^2}
789
+
790
+ This is the "wideband" transformation, producing a stopband with
791
+ geometric (log frequency) symmetry about `wo`.
792
+
793
+ """
794
+ z = cupy.atleast_1d(z)
795
+ p = cupy.atleast_1d(p)
796
+ wo = float(wo)
797
+ bw = float(bw)
798
+
799
+ degree = _relative_degree(z, p)
800
+
801
+ # Invert to a highpass filter with desired bandwidth
802
+ z_hp = (bw/2) / z
803
+ p_hp = (bw/2) / p
804
+
805
+ # Square root needs to produce complex result, not NaN
806
+ z_hp = z_hp.astype(complex)
807
+ p_hp = p_hp.astype(complex)
808
+
809
+ # Duplicate poles and zeros and shift from baseband to +wo and -wo
810
+ z_bs = cupy.concatenate((z_hp + cupy.sqrt(z_hp**2 - wo**2),
811
+ z_hp - cupy.sqrt(z_hp**2 - wo**2)))
812
+ p_bs = cupy.concatenate((p_hp + cupy.sqrt(p_hp**2 - wo**2),
813
+ p_hp - cupy.sqrt(p_hp**2 - wo**2)))
814
+
815
+ # Move any zeros that were at infinity to the center of the stopband
816
+ z_bs = cupy.append(z_bs, cupy.full(degree, +1j*wo))
817
+ z_bs = cupy.append(z_bs, cupy.full(degree, -1j*wo))
818
+
819
+ # Cancel out gain change caused by inversion
820
+ k_bs = k * cupy.real(cupy.prod(-z) / cupy.prod(-p))
821
+
822
+ return z_bs, p_bs, k_bs
823
+
824
+
825
+ def bilinear(b, a, fs=1.0):
826
+ r"""
827
+ Return a digital IIR filter from an analog one using a bilinear transform.
828
+
829
+ Transform a set of poles and zeros from the analog s-plane to the digital
830
+ z-plane using Tustin's method, which substitutes ``2*fs*(z-1) / (z+1)`` for
831
+ ``s``, maintaining the shape of the frequency response.
832
+
833
+ Parameters
834
+ ----------
835
+ b : array_like
836
+ Numerator of the analog filter transfer function.
837
+ a : array_like
838
+ Denominator of the analog filter transfer function.
839
+ fs : float
840
+ Sample rate, as ordinary frequency (e.g., hertz). No prewarping is
841
+ done in this function.
842
+
843
+ Returns
844
+ -------
845
+ b : ndarray
846
+ Numerator of the transformed digital filter transfer function.
847
+ a : ndarray
848
+ Denominator of the transformed digital filter transfer function.
849
+
850
+ See Also
851
+ --------
852
+ lp2lp, lp2hp, lp2bp, lp2bs
853
+ bilinear_zpk
854
+ scipy.signal.bilinear
855
+
856
+ """
857
+ fs = float(fs)
858
+ a, b = map(cupy.atleast_1d, (a, b))
859
+ D = a.shape[0] - 1
860
+ N = b.shape[0] - 1
861
+
862
+ M = max(N, D)
863
+ Np, Dp = M, M
864
+
865
+ bprime = cupy.empty(Np + 1, float)
866
+ aprime = cupy.empty(Dp + 1, float)
867
+
868
+ # XXX (ev-br): worth turning into a ufunc invocation? (loops are short)
869
+ for j in range(Dp + 1):
870
+ val = 0.0
871
+ for i in range(N + 1):
872
+ bNi = b[N - i] * (2 * fs)**i
873
+ for k in range(i + 1):
874
+ for s in range(M - i + 1):
875
+ if k + s == j:
876
+ val += comb(i, k) * comb(M - i, s) * bNi * (-1)**k
877
+ bprime[j] = cupy.real(val)
878
+
879
+ for j in range(Dp + 1):
880
+ val = 0.0
881
+ for i in range(D + 1):
882
+ aDi = a[D - i] * (2 * fs)**i
883
+ for k in range(i + 1):
884
+ for s in range(M - i + 1):
885
+ if k + s == j:
886
+ val += comb(i, k) * comb(M - i, s) * aDi * (-1)**k
887
+ aprime[j] = cupy.real(val)
888
+
889
+ return normalize(bprime, aprime)
890
+
891
+
892
+ def lp2lp(b, a, wo=1.0):
893
+ r"""
894
+ Transform a lowpass filter prototype to a different frequency.
895
+
896
+ Return an analog low-pass filter with cutoff frequency `wo`
897
+ from an analog low-pass filter prototype with unity cutoff frequency, in
898
+ transfer function ('ba') representation.
899
+
900
+ Parameters
901
+ ----------
902
+ b : array_like
903
+ Numerator polynomial coefficients.
904
+ a : array_like
905
+ Denominator polynomial coefficients.
906
+ wo : float
907
+ Desired cutoff, as angular frequency (e.g. rad/s).
908
+ Defaults to no change.
909
+
910
+ Returns
911
+ -------
912
+ b : array_like
913
+ Numerator polynomial coefficients of the transformed low-pass filter.
914
+ a : array_like
915
+ Denominator polynomial coefficients of the transformed low-pass filter.
916
+
917
+ See Also
918
+ --------
919
+ lp2hp, lp2bp, lp2bs, bilinear
920
+ lp2lp_zpk
921
+ scipy.signal.lp2lp
922
+
923
+ Notes
924
+ -----
925
+ This is derived from the s-plane substitution
926
+
927
+ .. math:: s \rightarrow \frac{s}{\omega_0}
928
+
929
+ """
930
+ a, b = map(cupy.atleast_1d, (a, b))
931
+ try:
932
+ wo = float(wo)
933
+ except TypeError:
934
+ wo = float(wo[0])
935
+ d = len(a)
936
+ n = len(b)
937
+ M = max(d, n)
938
+ pwo = wo ** cupy.arange(M - 1, -1, -1)
939
+ start1 = max((n - d, 0))
940
+ start2 = max((d - n, 0))
941
+ b = b * pwo[start1] / pwo[start2:]
942
+ a = a * pwo[start1] / pwo[start1:]
943
+ return normalize(b, a)
944
+
945
+
946
+ def lp2hp(b, a, wo=1.0):
947
+ r"""
948
+ Transform a lowpass filter prototype to a highpass filter.
949
+
950
+ Return an analog high-pass filter with cutoff frequency `wo`
951
+ from an analog low-pass filter prototype with unity cutoff frequency, in
952
+ transfer function ('ba') representation.
953
+
954
+ Parameters
955
+ ----------
956
+ b : array_like
957
+ Numerator polynomial coefficients.
958
+ a : array_like
959
+ Denominator polynomial coefficients.
960
+ wo : float
961
+ Desired cutoff, as angular frequency (e.g., rad/s).
962
+ Defaults to no change.
963
+
964
+ Returns
965
+ -------
966
+ b : array_like
967
+ Numerator polynomial coefficients of the transformed high-pass filter.
968
+ a : array_like
969
+ Denominator polynomial coefficients of the transformed high-pass
970
+ filter.
971
+
972
+ See Also
973
+ --------
974
+ lp2lp, lp2bp, lp2bs, bilinear
975
+ lp2hp_zpk
976
+ scipy.signal.lp2hp
977
+
978
+ Notes
979
+ -----
980
+ This is derived from the s-plane substitution
981
+
982
+ .. math:: s \rightarrow \frac{\omega_0}{s}
983
+
984
+ This maintains symmetry of the lowpass and highpass responses on a
985
+ logarithmic scale.
986
+ """
987
+ a, b = map(cupy.atleast_1d, (a, b))
988
+ try:
989
+ wo = float(wo)
990
+ except TypeError:
991
+ wo = float(wo[0])
992
+ d = len(a)
993
+ n = len(b)
994
+ if wo != 1:
995
+ pwo = wo ** cupy.arange(max(d, n))
996
+ else:
997
+ pwo = cupy.ones(max(d, n), b.dtype)
998
+ if d >= n:
999
+ outa = a[::-1] * pwo
1000
+ outb = cupy.resize(b, (d,))
1001
+ outb[n:] = 0.0
1002
+ outb[:n] = b[::-1] * pwo[:n]
1003
+ else:
1004
+ outb = b[::-1] * pwo
1005
+ outa = cupy.resize(a, (n,))
1006
+ outa[d:] = 0.0
1007
+ outa[:d] = a[::-1] * pwo[:d]
1008
+
1009
+ return normalize(outb, outa)
1010
+
1011
+
1012
+ def lp2bp(b, a, wo=1.0, bw=1.0):
1013
+ r"""
1014
+ Transform a lowpass filter prototype to a bandpass filter.
1015
+
1016
+ Return an analog band-pass filter with center frequency `wo` and
1017
+ bandwidth `bw` from an analog low-pass filter prototype with unity
1018
+ cutoff frequency, in transfer function ('ba') representation.
1019
+
1020
+ Parameters
1021
+ ----------
1022
+ b : array_like
1023
+ Numerator polynomial coefficients.
1024
+ a : array_like
1025
+ Denominator polynomial coefficients.
1026
+ wo : float
1027
+ Desired passband center, as angular frequency (e.g., rad/s).
1028
+ Defaults to no change.
1029
+ bw : float
1030
+ Desired passband width, as angular frequency (e.g., rad/s).
1031
+ Defaults to 1.
1032
+
1033
+ Returns
1034
+ -------
1035
+ b : array_like
1036
+ Numerator polynomial coefficients of the transformed band-pass filter.
1037
+ a : array_like
1038
+ Denominator polynomial coefficients of the transformed band-pass
1039
+ filter.
1040
+
1041
+ See Also
1042
+ --------
1043
+ lp2lp, lp2hp, lp2bs, bilinear
1044
+ lp2bp_zpk
1045
+ scipy.signal.lp2bp
1046
+
1047
+ Notes
1048
+ -----
1049
+ This is derived from the s-plane substitution
1050
+
1051
+ .. math:: s \rightarrow \frac{s^2 + {\omega_0}^2}{s \cdot \mathrm{BW}}
1052
+
1053
+ This is the "wideband" transformation, producing a passband with
1054
+ geometric (log frequency) symmetry about `wo`.
1055
+
1056
+ """
1057
+ a, b = map(cupy.atleast_1d, (a, b))
1058
+ D = len(a) - 1
1059
+ N = len(b) - 1
1060
+ artype = cupy.mintypecode((a.dtype, b.dtype))
1061
+ ma = max(N, D)
1062
+ Np = N + ma
1063
+ Dp = D + ma
1064
+ bprime = cupy.empty(Np + 1, artype)
1065
+ aprime = cupy.empty(Dp + 1, artype)
1066
+ wosq = wo * wo
1067
+ for j in range(Np + 1):
1068
+ val = 0.0
1069
+ for i in range(0, N + 1):
1070
+ for k in range(0, i + 1):
1071
+ if ma - i + 2 * k == j:
1072
+ val += comb(i, k) * b[N - i] * (wosq) ** (i - k) / bw ** i
1073
+ bprime[Np - j] = val
1074
+
1075
+ for j in range(Dp + 1):
1076
+ val = 0.0
1077
+ for i in range(0, D + 1):
1078
+ for k in range(0, i + 1):
1079
+ if ma - i + 2 * k == j:
1080
+ val += comb(i, k) * a[D - i] * (wosq) ** (i - k) / bw ** i
1081
+ aprime[Dp - j] = val
1082
+
1083
+ return normalize(bprime, aprime)
1084
+
1085
+
1086
+ def lp2bs(b, a, wo=1.0, bw=1.0):
1087
+ r"""
1088
+ Transform a lowpass filter prototype to a bandstop filter.
1089
+
1090
+ Return an analog band-stop filter with center frequency `wo` and
1091
+ bandwidth `bw` from an analog low-pass filter prototype with unity
1092
+ cutoff frequency, in transfer function ('ba') representation.
1093
+
1094
+ Parameters
1095
+ ----------
1096
+ b : array_like
1097
+ Numerator polynomial coefficients.
1098
+ a : array_like
1099
+ Denominator polynomial coefficients.
1100
+ wo : float
1101
+ Desired stopband center, as angular frequency (e.g., rad/s).
1102
+ Defaults to no change.
1103
+ bw : float
1104
+ Desired stopband width, as angular frequency (e.g., rad/s).
1105
+ Defaults to 1.
1106
+
1107
+ Returns
1108
+ -------
1109
+ b : array_like
1110
+ Numerator polynomial coefficients of the transformed band-stop filter.
1111
+ a : array_like
1112
+ Denominator polynomial coefficients of the transformed band-stop
1113
+ filter.
1114
+
1115
+ See Also
1116
+ --------
1117
+ lp2lp, lp2hp, lp2bp, bilinear
1118
+ lp2bs_zpk
1119
+ scipy.signal.lp2bs
1120
+
1121
+ Notes
1122
+ -----
1123
+ This is derived from the s-plane substitution
1124
+
1125
+ .. math:: s \rightarrow \frac{s \cdot \mathrm{BW}}{s^2 + {\omega_0}^2}
1126
+
1127
+ This is the "wideband" transformation, producing a stopband with
1128
+ geometric (log frequency) symmetry about `wo`.
1129
+ """
1130
+ a, b = map(cupy.atleast_1d, (a, b))
1131
+ D = len(a) - 1
1132
+ N = len(b) - 1
1133
+ artype = cupy.mintypecode((a.dtype, b.dtype))
1134
+ M = max(N, D)
1135
+ Np = M + M
1136
+ Dp = M + M
1137
+ bprime = cupy.empty(Np + 1, artype)
1138
+ aprime = cupy.empty(Dp + 1, artype)
1139
+ wosq = wo * wo
1140
+ for j in range(Np + 1):
1141
+ val = 0.0
1142
+ for i in range(0, N + 1):
1143
+ for k in range(0, M - i + 1):
1144
+ if i + 2 * k == j:
1145
+ val += (comb(M - i, k) * b[N - i] *
1146
+ (wosq) ** (M - i - k) * bw ** i)
1147
+ bprime[Np - j] = val
1148
+
1149
+ for j in range(Dp + 1):
1150
+ val = 0.0
1151
+ for i in range(0, D + 1):
1152
+ for k in range(0, M - i + 1):
1153
+ if i + 2 * k == j:
1154
+ val += (comb(M - i, k) * a[D - i] *
1155
+ (wosq) ** (M - i - k) * bw ** i)
1156
+ aprime[Dp - j] = val
1157
+
1158
+ return normalize(bprime, aprime)
1159
+
1160
+
1161
+ # ### LTI conversions ###
1162
+
1163
+ def zpk2tf(z, p, k):
1164
+ """
1165
+ Return polynomial transfer function representation from zeros and poles
1166
+
1167
+ Parameters
1168
+ ----------
1169
+ z : array_like
1170
+ Zeros of the transfer function.
1171
+ p : array_like
1172
+ Poles of the transfer function.
1173
+ k : float
1174
+ System gain.
1175
+
1176
+ Returns
1177
+ -------
1178
+ b : ndarray
1179
+ Numerator polynomial coefficients.
1180
+ a : ndarray
1181
+ Denominator polynomial coefficients.
1182
+
1183
+ See Also
1184
+ --------
1185
+ scipy.signal.zpk2tf
1186
+ """
1187
+ if z.ndim > 1:
1188
+ raise NotImplementedError(f"zpk2tf: z.ndim = {z.ndim}.")
1189
+ b = _polycoeffs_from_zeros(z) * k
1190
+ a = _polycoeffs_from_zeros(p)
1191
+ return b, a
1192
+
1193
+
1194
+ def tf2zpk(b, a):
1195
+ r"""Return zero, pole, gain (z, p, k) representation from a numerator,
1196
+ denominator representation of a linear filter.
1197
+
1198
+ Parameters
1199
+ ----------
1200
+ b : array_like
1201
+ Numerator polynomial coefficients.
1202
+ a : array_like
1203
+ Denominator polynomial coefficients.
1204
+
1205
+ Returns
1206
+ -------
1207
+ z : ndarray
1208
+ Zeros of the transfer function.
1209
+ p : ndarray
1210
+ Poles of the transfer function.
1211
+ k : float
1212
+ System gain.
1213
+
1214
+ Warning
1215
+ -------
1216
+ This function may synchronize the device.
1217
+
1218
+ See Also
1219
+ --------
1220
+ scipy.signal.tf2zpk
1221
+
1222
+ Notes
1223
+ -----
1224
+ If some values of `b` are too close to 0, they are removed. In that case,
1225
+ a BadCoefficients warning is emitted.
1226
+
1227
+ The `b` and `a` arrays are interpreted as coefficients for positive,
1228
+ descending powers of the transfer function variable. So the inputs
1229
+ :math:`b = [b_0, b_1, ..., b_M]` and :math:`a =[a_0, a_1, ..., a_N]`
1230
+ can represent an analog filter of the form:
1231
+
1232
+ .. math::
1233
+
1234
+ H(s) = \frac
1235
+ {b_0 s^M + b_1 s^{(M-1)} + \cdots + b_M}
1236
+ {a_0 s^N + a_1 s^{(N-1)} + \cdots + a_N}
1237
+
1238
+ or a discrete-time filter of the form:
1239
+
1240
+ .. math::
1241
+
1242
+ H(z) = \frac
1243
+ {b_0 z^M + b_1 z^{(M-1)} + \cdots + b_M}
1244
+ {a_0 z^N + a_1 z^{(N-1)} + \cdots + a_N}
1245
+
1246
+ This "positive powers" form is found more commonly in controls
1247
+ engineering. If `M` and `N` are equal (which is true for all filters
1248
+ generated by the bilinear transform), then this happens to be equivalent
1249
+ to the "negative powers" discrete-time form preferred in DSP:
1250
+
1251
+ .. math::
1252
+
1253
+ H(z) = \frac
1254
+ {b_0 + b_1 z^{-1} + \cdots + b_M z^{-M}}
1255
+ {a_0 + a_1 z^{-1} + \cdots + a_N z^{-N}}
1256
+
1257
+ Although this is true for common filters, remember that this is not true
1258
+ in the general case. If `M` and `N` are not equal, the discrete-time
1259
+ transfer function coefficients must first be converted to the "positive
1260
+ powers" form before finding the poles and zeros.
1261
+
1262
+ """
1263
+ b, a = normalize(b, a)
1264
+ b = (b + 0.0) / a[0]
1265
+ a = (a + 0.0) / a[0]
1266
+ k = b[0].copy()
1267
+ b /= b[0]
1268
+ z = roots(b)
1269
+ p = roots(a)
1270
+ return z, p, k
1271
+
1272
+
1273
+ def tf2sos(b, a, pairing=None, *, analog=False):
1274
+ """
1275
+ Return second-order sections from transfer function representation
1276
+
1277
+ Parameters
1278
+ ----------
1279
+ b : array_like
1280
+ Numerator polynomial coefficients.
1281
+ a : array_like
1282
+ Denominator polynomial coefficients.
1283
+ pairing : {None, 'nearest', 'keep_odd', 'minimal'}, optional
1284
+ The method to use to combine pairs of poles and zeros into sections.
1285
+ See `zpk2sos` for information and restrictions on `pairing` and
1286
+ `analog` arguments.
1287
+ analog : bool, optional
1288
+ If True, system is analog, otherwise discrete.
1289
+
1290
+ Returns
1291
+ -------
1292
+ sos : ndarray
1293
+ Array of second-order filter coefficients, with shape
1294
+ ``(n_sections, 6)``. See `sosfilt` for the SOS filter format
1295
+ specification.
1296
+
1297
+ See Also
1298
+ --------
1299
+ scipy.signal.tf2sos
1300
+
1301
+ Notes
1302
+ -----
1303
+ It is generally discouraged to convert from TF to SOS format, since doing
1304
+ so usually will not improve numerical precision errors. Instead, consider
1305
+ designing filters in ZPK format and converting directly to SOS. TF is
1306
+ converted to SOS by first converting to ZPK format, then converting
1307
+ ZPK to SOS.
1308
+
1309
+ """
1310
+ return zpk2sos(*tf2zpk(b, a), pairing=pairing, analog=analog)
1311
+
1312
+
1313
+ def sos2tf(sos):
1314
+ """
1315
+ Return a single transfer function from a series of second-order sections
1316
+
1317
+ Parameters
1318
+ ----------
1319
+ sos : array_like
1320
+ Array of second-order filter coefficients, must have shape
1321
+ ``(n_sections, 6)``. See `sosfilt` for the SOS filter format
1322
+ specification.
1323
+
1324
+ Returns
1325
+ -------
1326
+ b : ndarray
1327
+ Numerator polynomial coefficients.
1328
+ a : ndarray
1329
+ Denominator polynomial coefficients.
1330
+
1331
+ See Also
1332
+ --------
1333
+ scipy.signal.sos2tf
1334
+
1335
+ """
1336
+ sos = cupy.asarray(sos)
1337
+ result_type = sos.dtype
1338
+ if result_type.kind in 'bui':
1339
+ result_type = cupy.float64
1340
+
1341
+ b = cupy.array([1], dtype=result_type)
1342
+ a = cupy.array([1], dtype=result_type)
1343
+ n_sections = sos.shape[0]
1344
+ for section in range(n_sections):
1345
+ b = cupy.polymul(b, sos[section, :3])
1346
+ a = cupy.polymul(a, sos[section, 3:])
1347
+ return b, a
1348
+
1349
+
1350
+ def sos2zpk(sos):
1351
+ """
1352
+ Return zeros, poles, and gain of a series of second-order sections
1353
+
1354
+ Parameters
1355
+ ----------
1356
+ sos : array_like
1357
+ Array of second-order filter coefficients, must have shape
1358
+ ``(n_sections, 6)``. See `sosfilt` for the SOS filter format
1359
+ specification.
1360
+
1361
+ Returns
1362
+ -------
1363
+ z : ndarray
1364
+ Zeros of the transfer function.
1365
+ p : ndarray
1366
+ Poles of the transfer function.
1367
+ k : float
1368
+ System gain.
1369
+
1370
+ Notes
1371
+ -----
1372
+ The number of zeros and poles returned will be ``n_sections * 2``
1373
+ even if some of these are (effectively) zero.
1374
+
1375
+ See Also
1376
+ --------
1377
+ scipy.signal.sos2zpk
1378
+
1379
+ """
1380
+ n_sections = sos.shape[0]
1381
+ z = cupy.zeros(n_sections*2, cupy.complex128)
1382
+ p = cupy.zeros(n_sections*2, cupy.complex128)
1383
+ k = 1.
1384
+ for section in range(n_sections):
1385
+ # XXX: may just solve a quadratic equation instead of tf2zpk
1386
+ zpk = tf2zpk(sos[section, :3], sos[section, 3:])
1387
+ z[2*section:2*section + len(zpk[0])] = zpk[0]
1388
+ p[2*section:2*section + len(zpk[1])] = zpk[1]
1389
+ k *= zpk[2]
1390
+ return z, p, k
1391
+
1392
+
1393
+ def tf2ss(num, den):
1394
+ r"""Transfer function to state-space representation.
1395
+
1396
+ Parameters
1397
+ ----------
1398
+ num, den : array_like
1399
+ Sequences representing the coefficients of the numerator and
1400
+ denominator polynomials, in order of descending degree. The
1401
+ denominator needs to be at least as long as the numerator.
1402
+
1403
+ Returns
1404
+ -------
1405
+ A, B, C, D : ndarray
1406
+ State space representation of the system, in controller canonical
1407
+ form.
1408
+
1409
+ See Also
1410
+ --------
1411
+ scipy.signal.tf2ss
1412
+ """
1413
+ # Controller canonical state-space representation.
1414
+ # if M+1 = len(num) and K+1 = len(den) then we must have M <= K
1415
+ # states are found by asserting that X(s) = U(s) / D(s)
1416
+ # then Y(s) = N(s) * X(s)
1417
+ #
1418
+ # A, B, C, and D follow quite naturally.
1419
+ #
1420
+ num, den = normalize(num, den) # Strips zeros, checks arrays
1421
+ nn = len(num.shape)
1422
+ if nn == 1:
1423
+ num = cupy.asarray([num], num.dtype)
1424
+ M = num.shape[1]
1425
+ K = len(den)
1426
+ if M > K:
1427
+ msg = "Improper transfer function. `num` is longer than `den`."
1428
+ raise ValueError(msg)
1429
+ if M == 0 or K == 0: # Null system
1430
+ return (cupy.array([], float),
1431
+ cupy.array([], float),
1432
+ cupy.array([], float),
1433
+ cupy.array([], float))
1434
+
1435
+ # pad numerator to have same number of columns has denominator
1436
+ num = cupy.hstack((cupy.zeros((num.shape[0], K - M), num.dtype), num))
1437
+
1438
+ if num.shape[-1] > 0:
1439
+ D = cupy.atleast_2d(num[:, 0])
1440
+
1441
+ else:
1442
+ # We don't assign it an empty array because this system
1443
+ # is not 'null'. It just doesn't have a non-zero D
1444
+ # matrix. Thus, it should have a non-zero shape so that
1445
+ # it can be operated on by functions like 'ss2tf'
1446
+ D = cupy.array([[0]], float)
1447
+
1448
+ if K == 1:
1449
+ D = D.reshape(num.shape)
1450
+
1451
+ return (cupy.zeros((1, 1)), cupy.zeros((1, D.shape[1])),
1452
+ cupy.zeros((D.shape[0], 1)), D)
1453
+
1454
+ frow = -cupy.array([den[1:]])
1455
+ A = cupy.r_[frow, cupy.eye(K - 2, K - 1)]
1456
+ B = cupy.eye(K - 1, 1)
1457
+ C = num[:, 1:] - cupy.outer(num[:, 0], den[1:])
1458
+ D = D.reshape((C.shape[0], B.shape[1]))
1459
+
1460
+ return A, B, C, D
1461
+
1462
+
1463
+ def ss2tf(A, B, C, D, input=0):
1464
+ r"""State-space to transfer function.
1465
+
1466
+ A, B, C, D defines a linear state-space system with `p` inputs,
1467
+ `q` outputs, and `n` state variables.
1468
+
1469
+ Parameters
1470
+ ----------
1471
+ A : array_like
1472
+ State (or system) matrix of shape ``(n, n)``
1473
+ B : array_like
1474
+ Input matrix of shape ``(n, p)``
1475
+ C : array_like
1476
+ Output matrix of shape ``(q, n)``
1477
+ D : array_like
1478
+ Feedthrough (or feedforward) matrix of shape ``(q, p)``
1479
+ input : int, optional
1480
+ For multiple-input systems, the index of the input to use.
1481
+
1482
+ Returns
1483
+ -------
1484
+ num : 2-D ndarray
1485
+ Numerator(s) of the resulting transfer function(s). `num` has one row
1486
+ for each of the system's outputs. Each row is a sequence representation
1487
+ of the numerator polynomial.
1488
+ den : 1-D ndarray
1489
+ Denominator of the resulting transfer function(s). `den` is a sequence
1490
+ representation of the denominator polynomial.
1491
+
1492
+ Warning
1493
+ -------
1494
+ This function may synchronize the device.
1495
+
1496
+ See Also
1497
+ --------
1498
+ scipy.signal.ss2tf
1499
+
1500
+ """
1501
+ # transfer function is C (sI - A)**(-1) B + D
1502
+
1503
+ # Check consistency and make them all rank-2 arrays
1504
+ A, B, C, D = abcd_normalize(A, B, C, D)
1505
+
1506
+ nout, nin = D.shape
1507
+ if input >= nin:
1508
+ raise ValueError("System does not have the input specified.")
1509
+
1510
+ # make SIMO from possibly MIMO system.
1511
+ B = B[:, input:input + 1]
1512
+ D = D[:, input:input + 1]
1513
+
1514
+ try:
1515
+ den = poly(A)
1516
+ except ValueError:
1517
+ den = 1
1518
+
1519
+ if (prod(B.shape) == 0) and (prod(C.shape) == 0):
1520
+ num = cupy.ravel(D)
1521
+ if (prod(D.shape) == 0) and (prod(A.shape) == 0):
1522
+ den = []
1523
+ return num, den
1524
+
1525
+ num_states = A.shape[0]
1526
+ type_test = A[:, 0] + B[:, 0] + C[0, :] + D + 0.0
1527
+ num = cupy.empty((nout, num_states + 1), type_test.dtype)
1528
+ for k in range(nout):
1529
+ Ck = cupy.atleast_2d(C[k, :])
1530
+ num[k] = poly(A - B @ Ck) + (D[k] - 1) * den
1531
+
1532
+ return num, den
1533
+
1534
+
1535
+ def zpk2ss(z, p, k):
1536
+ """Zero-pole-gain representation to state-space representation
1537
+
1538
+ Parameters
1539
+ ----------
1540
+ z, p : sequence
1541
+ Zeros and poles.
1542
+ k : float
1543
+ System gain.
1544
+
1545
+ Returns
1546
+ -------
1547
+ A, B, C, D : ndarray
1548
+ State space representation of the system, in controller canonical
1549
+ form.
1550
+
1551
+ See Also
1552
+ --------
1553
+ scipy.signal.zpk2ss
1554
+
1555
+ """
1556
+ return tf2ss(*zpk2tf(z, p, k))
1557
+
1558
+
1559
+ def ss2zpk(A, B, C, D, input=0):
1560
+ """State-space representation to zero-pole-gain representation.
1561
+
1562
+ A, B, C, D defines a linear state-space system with `p` inputs,
1563
+ `q` outputs, and `n` state variables.
1564
+
1565
+ Parameters
1566
+ ----------
1567
+ A : array_like
1568
+ State (or system) matrix of shape ``(n, n)``
1569
+ B : array_like
1570
+ Input matrix of shape ``(n, p)``
1571
+ C : array_like
1572
+ Output matrix of shape ``(q, n)``
1573
+ D : array_like
1574
+ Feedthrough (or feedforward) matrix of shape ``(q, p)``
1575
+ input : int, optional
1576
+ For multiple-input systems, the index of the input to use.
1577
+
1578
+ Returns
1579
+ -------
1580
+ z, p : sequence
1581
+ Zeros and poles.
1582
+ k : float
1583
+ System gain.
1584
+
1585
+ See Also
1586
+ --------
1587
+ scipy.signal.ss2zpk
1588
+
1589
+ """
1590
+ return tf2zpk(*ss2tf(A, B, C, D, input=input))
1591
+
1592
+
1593
+ # ### Low-level analog filter prototypes ###
1594
+
1595
+ # TODO (ev-br): move to a better place (_filter_design.py (?))
1596
+
1597
+ def buttap(N):
1598
+ """Return (z,p,k) for analog prototype of Nth-order Butterworth filter.
1599
+
1600
+ The filter will have an angular (e.g., rad/s) cutoff frequency of 1.
1601
+
1602
+ See Also
1603
+ --------
1604
+ butter : Filter design function using this prototype
1605
+ scipy.signal.buttap
1606
+
1607
+ """
1608
+ if abs(int(N)) != N:
1609
+ raise ValueError("Filter order must be a nonnegative integer")
1610
+ z = cupy.array([])
1611
+ m = cupy.arange(-N+1, N, 2)
1612
+ # Middle value is 0 to ensure an exactly real pole
1613
+ p = -cupy.exp(1j * pi * m / (2 * N))
1614
+ k = 1
1615
+ return z, p, k
1616
+
1617
+
1618
+ def cheb1ap(N, rp):
1619
+ """
1620
+ Return (z,p,k) for Nth-order Chebyshev type I analog lowpass filter.
1621
+
1622
+ The returned filter prototype has `rp` decibels of ripple in the passband.
1623
+
1624
+ The filter's angular (e.g. rad/s) cutoff frequency is normalized to 1,
1625
+ defined as the point at which the gain first drops below ``-rp``.
1626
+
1627
+ See Also
1628
+ --------
1629
+ cheby1 : Filter design function using this prototype
1630
+
1631
+ """
1632
+ if abs(int(N)) != N:
1633
+ raise ValueError("Filter order must be a nonnegative integer")
1634
+ elif N == 0:
1635
+ # Avoid divide-by-zero error
1636
+ # Even order filters have DC gain of -rp dB
1637
+ return cupy.array([]), cupy.array([]), 10**(-rp/20)
1638
+ z = cupy.array([])
1639
+
1640
+ # Ripple factor (epsilon)
1641
+ eps = cupy.sqrt(10 ** (0.1 * rp) - 1.0)
1642
+ mu = 1.0 / N * cupy.arcsinh(1 / eps)
1643
+
1644
+ # Arrange poles in an ellipse on the left half of the S-plane
1645
+ m = cupy.arange(-N+1, N, 2)
1646
+ theta = pi * m / (2*N)
1647
+ p = -cupy.sinh(mu + 1j*theta)
1648
+
1649
+ k = cupy.prod(-p, axis=0).real
1650
+ if N % 2 == 0:
1651
+ k = k / cupy.sqrt(1 + eps * eps)
1652
+
1653
+ return z, p, k
1654
+
1655
+
1656
+ def cheb2ap(N, rs):
1657
+ """
1658
+ Return (z,p,k) for Nth-order Chebyshev type I analog lowpass filter.
1659
+
1660
+ The returned filter prototype has `rs` decibels of ripple in the stopband.
1661
+
1662
+ The filter's angular (e.g. rad/s) cutoff frequency is normalized to 1,
1663
+ defined as the point at which the gain first reaches ``-rs``.
1664
+
1665
+ See Also
1666
+ --------
1667
+ cheby2 : Filter design function using this prototype
1668
+
1669
+ """
1670
+ if abs(int(N)) != N:
1671
+ raise ValueError("Filter order must be a nonnegative integer")
1672
+ elif N == 0:
1673
+ # Avoid divide-by-zero warning
1674
+ return cupy.array([]), cupy.array([]), 1
1675
+
1676
+ # Ripple factor (epsilon)
1677
+ de = 1.0 / cupy.sqrt(10 ** (0.1 * rs) - 1)
1678
+ mu = cupy.arcsinh(1.0 / de) / N
1679
+
1680
+ if N % 2:
1681
+ m = cupy.concatenate((cupy.arange(-N+1, 0, 2),
1682
+ cupy.arange(2, N, 2)))
1683
+ else:
1684
+ m = cupy.arange(-N+1, N, 2)
1685
+
1686
+ z = -cupy.conjugate(1j / cupy.sin(m * pi / (2.0 * N)))
1687
+
1688
+ # Poles around the unit circle like Butterworth
1689
+ p = -cupy.exp(1j * pi * cupy.arange(-N+1, N, 2) / (2 * N))
1690
+ # Warp into Chebyshev II
1691
+ p = cupy.sinh(mu) * p.real + 1j * cupy.cosh(mu) * p.imag
1692
+ p = 1.0 / p
1693
+
1694
+ k = (cupy.prod(-p, axis=0) / cupy.prod(-z, axis=0)).real
1695
+ return z, p, k
1696
+
1697
+
1698
+ # ### Elliptic filter prototype ###
1699
+
1700
+ _POW10_LOG10 = math.log(10)
1701
+
1702
+
1703
+ def _pow10m1(x):
1704
+ """10 ** x - 1 for x near 0"""
1705
+ return cupy.expm1(_POW10_LOG10 * x)
1706
+
1707
+
1708
+ def _ellipdeg(n, m1):
1709
+ """Solve degree equation using nomes
1710
+
1711
+ Given n, m1, solve
1712
+ n * K(m) / K'(m) = K1(m1) / K1'(m1)
1713
+ for m
1714
+
1715
+ See [1], Eq. (49)
1716
+
1717
+ References
1718
+ ----------
1719
+ .. [1] Orfanidis, "Lecture Notes on Elliptic Filter Design",
1720
+ https://www.ece.rutgers.edu/~orfanidi/ece521/notes.pdf
1721
+ """
1722
+ # number of terms in solving degree equation
1723
+ _ELLIPDEG_MMAX = 7
1724
+
1725
+ K1 = special.ellipk(m1)
1726
+ K1p = special.ellipkm1(m1)
1727
+
1728
+ q1 = cupy.exp(-pi * K1p / K1)
1729
+ q = q1 ** (1/n)
1730
+
1731
+ mnum = cupy.arange(_ELLIPDEG_MMAX + 1)
1732
+ mden = cupy.arange(1, _ELLIPDEG_MMAX + 2)
1733
+
1734
+ num = (q ** (mnum * (mnum+1))).sum()
1735
+ den = 1 + 2 * (q ** (mden**2)).sum()
1736
+
1737
+ return 16 * q * (num / den) ** 4
1738
+
1739
+
1740
+ def _arc_jac_sn(w, m):
1741
+ """Inverse Jacobian elliptic sn
1742
+
1743
+ Solve for z in w = sn(z, m)
1744
+
1745
+ Parameters
1746
+ ----------
1747
+ w : complex scalar
1748
+ argument
1749
+
1750
+ m : scalar
1751
+ modulus; in interval [0, 1]
1752
+
1753
+
1754
+ See [1], Eq. (56)
1755
+
1756
+ References
1757
+ ----------
1758
+ .. [1] Orfanidis, "Lecture Notes on Elliptic Filter Design",
1759
+ https://www.ece.rutgers.edu/~orfanidi/ece521/notes.pdf
1760
+
1761
+ """
1762
+ # Maximum number of iterations in Landen transformation recursion
1763
+ # sequence. 10 is conservative; unit tests pass with 4, Orfanidis
1764
+ # (see _arc_jac_cn [1]) suggests 5.
1765
+ _ARC_JAC_SN_MAXITER = 10
1766
+
1767
+ def _complement(kx):
1768
+ # (1-k**2) ** 0.5; the expression below
1769
+ # works for small kx
1770
+ return ((1 - kx) * (1 + kx)) ** 0.5
1771
+
1772
+ k = m ** 0.5
1773
+
1774
+ if k > 1:
1775
+ return cupy.nan
1776
+ elif k == 1:
1777
+ return cupy.arctanh(w)
1778
+
1779
+ ks = [k]
1780
+ niter = 0
1781
+ while ks[-1] != 0:
1782
+ k_ = ks[-1]
1783
+ k_p = _complement(k_)
1784
+ ks.append((1 - k_p) / (1 + k_p))
1785
+ niter += 1
1786
+ if niter > _ARC_JAC_SN_MAXITER:
1787
+ raise ValueError('Landen transformation not converging')
1788
+
1789
+ K = cupy.prod(1 + cupy.array(ks[1:])) * pi/2
1790
+
1791
+ wns = [w]
1792
+
1793
+ for kn, knext in zip(ks[:-1], ks[1:]):
1794
+ wn = wns[-1]
1795
+ wnext = (2 * wn /
1796
+ ((1 + knext) * (1 + _complement(kn * wn))))
1797
+ wns.append(wnext)
1798
+
1799
+ u = 2 / pi * cupy.arcsin(wns[-1])
1800
+
1801
+ z = K * u
1802
+ return z
1803
+
1804
+
1805
+ def _arc_jac_sc1(w, m):
1806
+ """Real inverse Jacobian sc, with complementary modulus
1807
+
1808
+ Solve for z in w = sc(z, 1-m)
1809
+
1810
+ w - real scalar
1811
+
1812
+ m - modulus
1813
+
1814
+ Using that sc(z, m) = -i * sn(i * z, 1 - m)
1815
+ cf scipy/signal/_filter_design.py analog for an explanation
1816
+ and a reference.
1817
+
1818
+ """
1819
+
1820
+ zcomplex = _arc_jac_sn(1j * w, m)
1821
+ if abs(zcomplex.real) > 1e-14:
1822
+ raise ValueError
1823
+
1824
+ return zcomplex.imag
1825
+
1826
+
1827
+ def ellipap(N, rp, rs):
1828
+ """Return (z,p,k) of Nth-order elliptic analog lowpass filter.
1829
+
1830
+ The filter is a normalized prototype that has `rp` decibels of ripple
1831
+ in the passband and a stopband `rs` decibels down.
1832
+
1833
+ The filter's angular (e.g., rad/s) cutoff frequency is normalized to 1,
1834
+ defined as the point at which the gain first drops below ``-rp``.
1835
+
1836
+ See Also
1837
+ --------
1838
+ ellip : Filter design function using this prototype
1839
+ scipy.signal.elliap
1840
+
1841
+ """
1842
+ if abs(int(N)) != N:
1843
+ raise ValueError("Filter order must be a nonnegative integer")
1844
+ elif N == 0:
1845
+ # Avoid divide-by-zero warning
1846
+ # Even order filters have DC gain of -rp dB
1847
+ return cupy.array([]), cupy.array([]), 10**(-rp/20)
1848
+ elif N == 1:
1849
+ p = -cupy.sqrt(1.0 / _pow10m1(0.1 * rp))
1850
+ k = -p
1851
+ z = []
1852
+ return cupy.asarray(z), cupy.asarray(p), k
1853
+
1854
+ eps_sq = _pow10m1(0.1 * rp)
1855
+
1856
+ eps = cupy.sqrt(eps_sq)
1857
+ ck1_sq = eps_sq / _pow10m1(0.1 * rs)
1858
+ if ck1_sq == 0:
1859
+ raise ValueError("Cannot design a filter with given rp and rs"
1860
+ " specifications.")
1861
+
1862
+ m = _ellipdeg(N, ck1_sq)
1863
+ capk = special.ellipk(m)
1864
+ j = cupy.arange(1 - N % 2, N, 2)
1865
+ EPSILON = 2e-16
1866
+
1867
+ s, c, d, phi = special.ellipj(j * capk / N, m * cupy.ones_like(j))
1868
+ snew = cupy.compress(cupy.abs(s) > EPSILON, s, axis=-1)
1869
+ z = 1.j / (cupy.sqrt(m) * snew)
1870
+ z = cupy.concatenate((z, z.conj()))
1871
+
1872
+ r = _arc_jac_sc1(1. / eps, ck1_sq)
1873
+ v0 = capk * r / (N * special.ellipk(ck1_sq))
1874
+
1875
+ sv, cv, dv, phi = special.ellipj(v0, 1 - m)
1876
+ p = -(c * d * sv * cv + 1j * s * dv) / (1 - (d * sv) ** 2.0)
1877
+
1878
+ if N % 2:
1879
+ mask = cupy.abs(p.imag) > EPSILON * \
1880
+ cupy.sqrt((p * p.conj()).sum(axis=0).real)
1881
+ newp = cupy.compress(mask, p, axis=-1)
1882
+ p = cupy.concatenate((p, newp.conj()))
1883
+ else:
1884
+ p = cupy.concatenate((p, p.conj()))
1885
+
1886
+ k = (cupy.prod(-p, axis=0) / cupy.prod(-z, axis=0)).real
1887
+ if N % 2 == 0:
1888
+ k = k / cupy.sqrt(1 + eps_sq)
1889
+
1890
+ return z, p, k
1891
+
1892
+
1893
+ # ### *ord functions to accopany *ap functions
1894
+
1895
+ def _validate_gpass_gstop(gpass, gstop):
1896
+
1897
+ if gpass <= 0.0:
1898
+ raise ValueError("gpass should be larger than 0.0")
1899
+ elif gstop <= 0.0:
1900
+ raise ValueError("gstop should be larger than 0.0")
1901
+ elif gpass > gstop:
1902
+ raise ValueError("gpass should be smaller than gstop")
1903
+
1904
+
1905
+ def _pre_warp(wp, ws, analog):
1906
+ # Pre-warp frequencies for digital filter design
1907
+ if not analog:
1908
+ passb = cupy.tan(pi * wp / 2.0)
1909
+ stopb = cupy.tan(pi * ws / 2.0)
1910
+ else:
1911
+ passb = wp * 1.0
1912
+ stopb = ws * 1.0
1913
+ return passb, stopb
1914
+
1915
+
1916
+ def _validate_wp_ws(wp, ws, fs, analog):
1917
+ wp = cupy.atleast_1d(wp)
1918
+ ws = cupy.atleast_1d(ws)
1919
+ if fs is not None:
1920
+ if analog:
1921
+ raise ValueError("fs cannot be specified for an analog filter")
1922
+ wp = 2 * wp / fs
1923
+ ws = 2 * ws / fs
1924
+
1925
+ filter_type = 2 * (len(wp) - 1) + 1
1926
+ if wp[0] >= ws[0]:
1927
+ filter_type += 1
1928
+
1929
+ return wp, ws, filter_type
1930
+
1931
+
1932
+ def _find_nat_freq(stopb, passb, gpass, gstop, filter_type, filter_kind):
1933
+ if filter_type == 1: # low
1934
+ nat = stopb / passb
1935
+ elif filter_type == 2: # high
1936
+ nat = passb / stopb
1937
+ elif filter_type == 3: # stop
1938
+ wp0 = _optimize.fminbound(band_stop_obj, passb[0], stopb[0] - 1e-12,
1939
+ args=(0, passb, stopb, gpass, gstop,
1940
+ filter_kind),
1941
+ disp=0)
1942
+ passb[0] = wp0
1943
+ wp1 = _optimize.fminbound(band_stop_obj, stopb[1] + 1e-12, passb[1],
1944
+ args=(1, passb, stopb, gpass, gstop,
1945
+ filter_kind),
1946
+ disp=0)
1947
+ passb[1] = wp1
1948
+ nat = ((stopb * (passb[0] - passb[1])) /
1949
+ (stopb ** 2 - passb[0] * passb[1]))
1950
+ elif filter_type == 4: # pass
1951
+ nat = ((stopb ** 2 - passb[0] * passb[1]) /
1952
+ (stopb * (passb[0] - passb[1])))
1953
+ else:
1954
+ raise ValueError(f"should not happen: {filter_type=}.")
1955
+
1956
+ nat = min(cupy.abs(nat))
1957
+ return nat, passb
1958
+
1959
+
1960
+ def _postprocess_wn(WN, analog, fs):
1961
+ wn = WN if analog else cupy.arctan(WN) * 2.0 / pi
1962
+ if len(wn) == 1:
1963
+ wn = wn[0]
1964
+ if fs is not None:
1965
+ wn = wn * fs / 2
1966
+ return wn
1967
+
1968
+
1969
+ def band_stop_obj(wp, ind, passb, stopb, gpass, gstop, type):
1970
+ """
1971
+ Band Stop Objective Function for order minimization.
1972
+
1973
+ Returns the non-integer order for an analog band stop filter.
1974
+
1975
+ Parameters
1976
+ ----------
1977
+ wp : scalar
1978
+ Edge of passband `passb`.
1979
+ ind : int, {0, 1}
1980
+ Index specifying which `passb` edge to vary (0 or 1).
1981
+ passb : ndarray
1982
+ Two element sequence of fixed passband edges.
1983
+ stopb : ndarray
1984
+ Two element sequence of fixed stopband edges.
1985
+ gstop : float
1986
+ Amount of attenuation in stopband in dB.
1987
+ gpass : float
1988
+ Amount of ripple in the passband in dB.
1989
+ type : {'butter', 'cheby', 'ellip'}
1990
+ Type of filter.
1991
+
1992
+ Returns
1993
+ -------
1994
+ n : scalar
1995
+ Filter order (possibly non-integer).
1996
+
1997
+ See Also
1998
+ --------
1999
+ scipy.signal.band_stop_obj
2000
+
2001
+ """
2002
+
2003
+ _validate_gpass_gstop(gpass, gstop)
2004
+
2005
+ passbC = passb.copy()
2006
+ passbC[ind] = wp
2007
+ nat = (stopb * (passbC[0] - passbC[1]) /
2008
+ (stopb ** 2 - passbC[0] * passbC[1]))
2009
+ nat = min(cupy.abs(nat))
2010
+
2011
+ if type == 'butter':
2012
+ GSTOP = 10 ** (0.1 * cupy.abs(gstop))
2013
+ GPASS = 10 ** (0.1 * cupy.abs(gpass))
2014
+ n = (cupy.log10((GSTOP - 1.0) / (GPASS - 1.0)) / (2 * cupy.log10(nat)))
2015
+ elif type == 'cheby':
2016
+ GSTOP = 10 ** (0.1 * cupy.abs(gstop))
2017
+ GPASS = 10 ** (0.1 * cupy.abs(gpass))
2018
+ n = cupy.arccosh(
2019
+ cupy.sqrt((GSTOP - 1.0) / (GPASS - 1.0))) / cupy.arccosh(nat)
2020
+ elif type == 'ellip':
2021
+ GSTOP = 10 ** (0.1 * gstop)
2022
+ GPASS = 10 ** (0.1 * gpass)
2023
+ arg1 = cupy.sqrt((GPASS - 1.0) / (GSTOP - 1.0))
2024
+ arg0 = 1.0 / nat
2025
+ d0 = special.ellipk(cupy.array([arg0 ** 2, 1 - arg0 ** 2]))
2026
+ d1 = special.ellipk(cupy.array([arg1 ** 2, 1 - arg1 ** 2]))
2027
+ n = (d0[0] * d1[1] / (d0[1] * d1[0]))
2028
+ else:
2029
+ raise ValueError("Incorrect type: %s" % type)
2030
+ return n
2031
+
2032
+
2033
+ def buttord(wp, ws, gpass, gstop, analog=False, fs=None):
2034
+ """Butterworth filter order selection.
2035
+
2036
+ Return the order of the lowest order digital or analog Butterworth filter
2037
+ that loses no more than `gpass` dB in the passband and has at least
2038
+ `gstop` dB attenuation in the stopband.
2039
+
2040
+ Parameters
2041
+ ----------
2042
+ wp, ws : float
2043
+ Passband and stopband edge frequencies.
2044
+
2045
+ For digital filters, these are in the same units as `fs`. By default,
2046
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
2047
+ where 1 is the Nyquist frequency. (`wp` and `ws` are thus in
2048
+ half-cycles / sample.) For example:
2049
+
2050
+ - Lowpass: wp = 0.2, ws = 0.3
2051
+ - Highpass: wp = 0.3, ws = 0.2
2052
+ - Bandpass: wp = [0.2, 0.5], ws = [0.1, 0.6]
2053
+ - Bandstop: wp = [0.1, 0.6], ws = [0.2, 0.5]
2054
+
2055
+ For analog filters, `wp` and `ws` are angular frequencies
2056
+ (e.g., rad/s).
2057
+ gpass : float
2058
+ The maximum loss in the passband (dB).
2059
+ gstop : float
2060
+ The minimum attenuation in the stopband (dB).
2061
+ analog : bool, optional
2062
+ When True, return an analog filter, otherwise a digital filter is
2063
+ returned.
2064
+ fs : float, optional
2065
+ The sampling frequency of the digital system.
2066
+
2067
+ .. versionadded:: 1.2.0
2068
+
2069
+ Returns
2070
+ -------
2071
+ ord : int
2072
+ The lowest order for a Butterworth filter which meets specs.
2073
+ wn : ndarray or float
2074
+ The Butterworth natural frequency (i.e. the "3dB frequency"). Should
2075
+ be used with `butter` to give filter results. If `fs` is specified,
2076
+ this is in the same units, and `fs` must also be passed to `butter`.
2077
+
2078
+ See Also
2079
+ --------
2080
+ scipy.signal.buttord
2081
+ butter : Filter design using order and critical points
2082
+ cheb1ord : Find order and critical points from passband and stopband spec
2083
+ cheb2ord, ellipord
2084
+ iirfilter : General filter design using order and critical frequencies
2085
+ iirdesign : General filter design using passband and stopband spec
2086
+
2087
+ """
2088
+ _validate_gpass_gstop(gpass, gstop)
2089
+ wp, ws, filter_type = _validate_wp_ws(wp, ws, fs, analog)
2090
+ passb, stopb = _pre_warp(wp, ws, analog)
2091
+ nat, passb = _find_nat_freq(
2092
+ stopb, passb, gpass, gstop, filter_type, 'butter')
2093
+
2094
+ GSTOP = 10 ** (0.1 * cupy.abs(gstop))
2095
+ GPASS = 10 ** (0.1 * cupy.abs(gpass))
2096
+ ord = int(cupy.ceil(cupy.log10((GSTOP - 1.0) /
2097
+ (GPASS - 1.0)) / (2 * cupy.log10(nat))))
2098
+
2099
+ # Find the Butterworth natural frequency WN (or the "3dB" frequency")
2100
+ # to give exactly gpass at passb.
2101
+ try:
2102
+ W0 = (GPASS - 1.0) ** (-1.0 / (2.0 * ord))
2103
+ except ZeroDivisionError:
2104
+ W0 = 1.0
2105
+ warnings.warn("Order is zero...check input parameters.",
2106
+ RuntimeWarning, 2)
2107
+
2108
+ # now convert this frequency back from lowpass prototype
2109
+ # to the original analog filter
2110
+
2111
+ if filter_type == 1: # low
2112
+ WN = W0 * passb
2113
+ elif filter_type == 2: # high
2114
+ WN = passb / W0
2115
+ elif filter_type == 3: # stop
2116
+ WN = cupy.empty(2, float)
2117
+ discr = cupy.sqrt((passb[1] - passb[0]) ** 2 +
2118
+ 4 * W0 ** 2 * passb[0] * passb[1])
2119
+ WN[0] = ((passb[1] - passb[0]) + discr) / (2 * W0)
2120
+ WN[1] = ((passb[1] - passb[0]) - discr) / (2 * W0)
2121
+ WN = cupy.sort(cupy.abs(WN))
2122
+ elif filter_type == 4: # pass
2123
+ W0 = cupy.array([-W0, W0], dtype=float)
2124
+ WN = (-W0 * (passb[1] - passb[0]) / 2.0 +
2125
+ cupy.sqrt(W0 ** 2 / 4.0 * (passb[1] - passb[0]) ** 2 +
2126
+ passb[0] * passb[1]))
2127
+ WN = cupy.sort(cupy.abs(WN))
2128
+ else:
2129
+ raise ValueError("Bad type: %s" % filter_type)
2130
+
2131
+ wn = _postprocess_wn(WN, analog, fs)
2132
+
2133
+ return ord, wn
2134
+
2135
+
2136
+ def cheb1ord(wp, ws, gpass, gstop, analog=False, fs=None):
2137
+ """Chebyshev type I filter order selection.
2138
+
2139
+ Return the order of the lowest order digital or analog Chebyshev Type I
2140
+ filter that loses no more than `gpass` dB in the passband and has at
2141
+ least `gstop` dB attenuation in the stopband.
2142
+
2143
+ Parameters
2144
+ ----------
2145
+ wp, ws : float
2146
+ Passband and stopband edge frequencies.
2147
+
2148
+ For digital filters, these are in the same units as `fs`. By default,
2149
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
2150
+ where 1 is the Nyquist frequency. (`wp` and `ws` are thus in
2151
+ half-cycles / sample.) For example:
2152
+
2153
+ - Lowpass: wp = 0.2, ws = 0.3
2154
+ - Highpass: wp = 0.3, ws = 0.2
2155
+ - Bandpass: wp = [0.2, 0.5], ws = [0.1, 0.6]
2156
+ - Bandstop: wp = [0.1, 0.6], ws = [0.2, 0.5]
2157
+
2158
+ For analog filters, `wp` and `ws` are angular frequencies
2159
+ (e.g., rad/s).
2160
+ gpass : float
2161
+ The maximum loss in the passband (dB).
2162
+ gstop : float
2163
+ The minimum attenuation in the stopband (dB).
2164
+ analog : bool, optional
2165
+ When True, return an analog filter, otherwise a digital filter is
2166
+ returned.
2167
+ fs : float, optional
2168
+ The sampling frequency of the digital system.
2169
+
2170
+ Returns
2171
+ -------
2172
+ ord : int
2173
+ The lowest order for a Chebyshev type I filter that meets specs.
2174
+ wn : ndarray or float
2175
+ The Chebyshev natural frequency (the "3dB frequency") for use with
2176
+ `cheby1` to give filter results. If `fs` is specified,
2177
+ this is in the same units, and `fs` must also be passed to `cheby1`.
2178
+
2179
+ See Also
2180
+ --------
2181
+ scipy.signal.cheb1ord
2182
+ cheby1 : Filter design using order and critical points
2183
+ buttord : Find order and critical points from passband and stopband spec
2184
+ cheb2ord, ellipord
2185
+ iirfilter : General filter design using order and critical frequencies
2186
+ iirdesign : General filter design using passband and stopband spec
2187
+
2188
+ """
2189
+ _validate_gpass_gstop(gpass, gstop)
2190
+ wp, ws, filter_type = _validate_wp_ws(wp, ws, fs, analog)
2191
+ passb, stopb = _pre_warp(wp, ws, analog)
2192
+ nat, passb = _find_nat_freq(
2193
+ stopb, passb, gpass, gstop, filter_type, 'cheby')
2194
+
2195
+ GSTOP = 10 ** (0.1 * cupy.abs(gstop))
2196
+ GPASS = 10 ** (0.1 * cupy.abs(gpass))
2197
+ v_pass_stop = cupy.arccosh(cupy.sqrt((GSTOP - 1.0) / (GPASS - 1.0)))
2198
+ ord = int(cupy.ceil(v_pass_stop / cupy.arccosh(nat)))
2199
+
2200
+ # Natural frequencies are just the passband edges
2201
+ wn = _postprocess_wn(passb, analog, fs)
2202
+
2203
+ return ord, wn
2204
+
2205
+
2206
+ def cheb2ord(wp, ws, gpass, gstop, analog=False, fs=None):
2207
+ """Chebyshev type II filter order selection.
2208
+
2209
+ Return the order of the lowest order digital or analog Chebyshev Type II
2210
+ filter that loses no more than `gpass` dB in the passband and has at least
2211
+ `gstop` dB attenuation in the stopband.
2212
+
2213
+ Parameters
2214
+ ----------
2215
+ wp, ws : float
2216
+ Passband and stopband edge frequencies.
2217
+
2218
+ For digital filters, these are in the same units as `fs`. By default,
2219
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
2220
+ where 1 is the Nyquist frequency. (`wp` and `ws` are thus in
2221
+ half-cycles / sample.) For example:
2222
+
2223
+ - Lowpass: wp = 0.2, ws = 0.3
2224
+ - Highpass: wp = 0.3, ws = 0.2
2225
+ - Bandpass: wp = [0.2, 0.5], ws = [0.1, 0.6]
2226
+ - Bandstop: wp = [0.1, 0.6], ws = [0.2, 0.5]
2227
+
2228
+ For analog filters, `wp` and `ws` are angular frequencies
2229
+ (e.g., rad/s).
2230
+ gpass : float
2231
+ The maximum loss in the passband (dB).
2232
+ gstop : float
2233
+ The minimum attenuation in the stopband (dB).
2234
+ analog : bool, optional
2235
+ When True, return an analog filter, otherwise a digital filter is
2236
+ returned.
2237
+ fs : float, optional
2238
+ The sampling frequency of the digital system.
2239
+
2240
+ Returns
2241
+ -------
2242
+ ord : int
2243
+ The lowest order for a Chebyshev type II filter that meets specs.
2244
+ wn : ndarray or float
2245
+ The Chebyshev natural frequency (the "3dB frequency") for use with
2246
+ `cheby2` to give filter results. If `fs` is specified,
2247
+ this is in the same units, and `fs` must also be passed to `cheby2`.
2248
+
2249
+ See Also
2250
+ --------
2251
+ scipy.signal.cheb2ord
2252
+ cheby2 : Filter design using order and critical points
2253
+ buttord : Find order and critical points from passband and stopband spec
2254
+ cheb1ord, ellipord
2255
+ iirfilter : General filter design using order and critical frequencies
2256
+ iirdesign : General filter design using passband and stopband spec
2257
+
2258
+ """
2259
+ _validate_gpass_gstop(gpass, gstop)
2260
+ wp, ws, filter_type = _validate_wp_ws(wp, ws, fs, analog)
2261
+ passb, stopb = _pre_warp(wp, ws, analog)
2262
+ nat, passb = _find_nat_freq(
2263
+ stopb, passb, gpass, gstop, filter_type, 'cheby')
2264
+
2265
+ GSTOP = 10 ** (0.1 * cupy.abs(gstop))
2266
+ GPASS = 10 ** (0.1 * cupy.abs(gpass))
2267
+ v_pass_stop = cupy.arccosh(cupy.sqrt((GSTOP - 1.0) / (GPASS - 1.0)))
2268
+ ord = int(cupy.ceil(v_pass_stop / cupy.arccosh(nat)))
2269
+
2270
+ # Find frequency where analog response is -gpass dB.
2271
+ # Then convert back from low-pass prototype to the original filter.
2272
+
2273
+ new_freq = cupy.cosh(1.0 / ord * v_pass_stop)
2274
+ new_freq = 1.0 / new_freq
2275
+
2276
+ if filter_type == 1:
2277
+ nat = passb / new_freq
2278
+ elif filter_type == 2:
2279
+ nat = passb * new_freq
2280
+ elif filter_type == 3:
2281
+ nat = cupy.empty(2, dtype=float)
2282
+ nat[0] = (new_freq / 2.0 * (passb[0] - passb[1]) +
2283
+ cupy.sqrt(new_freq ** 2 * (passb[1] - passb[0]) ** 2 / 4.0 +
2284
+ passb[1] * passb[0]))
2285
+ nat[1] = passb[1] * passb[0] / nat[0]
2286
+ elif filter_type == 4:
2287
+ nat = cupy.empty(2, dtype=float)
2288
+ nat[0] = (1.0 / (2.0 * new_freq) * (passb[0] - passb[1]) +
2289
+ cupy.sqrt((passb[1] - passb[0]) ** 2 / (4.0 * new_freq ** 2)
2290
+ + passb[1] * passb[0]))
2291
+ nat[1] = passb[0] * passb[1] / nat[0]
2292
+
2293
+ wn = _postprocess_wn(nat, analog, fs)
2294
+
2295
+ return ord, wn
2296
+
2297
+
2298
+ def ellipord(wp, ws, gpass, gstop, analog=False, fs=None):
2299
+ """Elliptic (Cauer) filter order selection.
2300
+
2301
+ Return the order of the lowest order digital or analog elliptic filter
2302
+ that loses no more than `gpass` dB in the passband and has at least
2303
+ `gstop` dB attenuation in the stopband.
2304
+
2305
+ Parameters
2306
+ ----------
2307
+ wp, ws : float
2308
+ Passband and stopband edge frequencies.
2309
+
2310
+ For digital filters, these are in the same units as `fs`. By default,
2311
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
2312
+ where 1 is the Nyquist frequency. (`wp` and `ws` are thus in
2313
+ half-cycles / sample.) For example:
2314
+
2315
+ - Lowpass: wp = 0.2, ws = 0.3
2316
+ - Highpass: wp = 0.3, ws = 0.2
2317
+ - Bandpass: wp = [0.2, 0.5], ws = [0.1, 0.6]
2318
+ - Bandstop: wp = [0.1, 0.6], ws = [0.2, 0.5]
2319
+
2320
+ For analog filters, `wp` and `ws` are angular frequencies
2321
+ (e.g., rad/s).
2322
+ gpass : float
2323
+ The maximum loss in the passband (dB).
2324
+ gstop : float
2325
+ The minimum attenuation in the stopband (dB).
2326
+ analog : bool, optional
2327
+ When True, return an analog filter, otherwise a digital filter is
2328
+ returned.
2329
+ fs : float, optional
2330
+ The sampling frequency of the digital system.
2331
+
2332
+ Returns
2333
+ -------
2334
+ ord : int
2335
+ The lowest order for an Elliptic (Cauer) filter that meets specs.
2336
+ wn : ndarray or float
2337
+ The Chebyshev natural frequency (the "3dB frequency") for use with
2338
+ `ellip` to give filter results. If `fs` is specified,
2339
+ this is in the same units, and `fs` must also be passed to `ellip`.
2340
+
2341
+ See Also
2342
+ --------
2343
+ scipy.signal.ellipord
2344
+ ellip : Filter design using order and critical points
2345
+ buttord : Find order and critical points from passband and stopband spec
2346
+ cheb1ord, cheb2ord
2347
+ iirfilter : General filter design using order and critical frequencies
2348
+ iirdesign : General filter design using passband and stopband spec
2349
+ """
2350
+ _validate_gpass_gstop(gpass, gstop)
2351
+ wp, ws, filter_type = _validate_wp_ws(wp, ws, fs, analog)
2352
+ passb, stopb = _pre_warp(wp, ws, analog)
2353
+ nat, passb = _find_nat_freq(
2354
+ stopb, passb, gpass, gstop, filter_type, 'ellip')
2355
+
2356
+ arg1_sq = _pow10m1(0.1 * gpass) / _pow10m1(0.1 * gstop)
2357
+ arg0 = 1.0 / nat
2358
+ d0 = special.ellipk(arg0 ** 2), special.ellipkm1(arg0 ** 2)
2359
+ d1 = special.ellipk(arg1_sq), special.ellipkm1(arg1_sq)
2360
+ ord = int(cupy.ceil(d0[0] * d1[1] / (d0[1] * d1[0])))
2361
+
2362
+ wn = _postprocess_wn(passb, analog, fs)
2363
+
2364
+ return ord, wn
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_iir_filter_design.py ADDED
@@ -0,0 +1,994 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """IIR filter design APIs"""
2
+ from math import pi
3
+ import math
4
+
5
+ import cupy
6
+
7
+ from cupyx.scipy.signal._iir_filter_conversions import (
8
+ lp2bp_zpk, lp2lp_zpk, lp2hp_zpk, lp2bs_zpk, bilinear_zpk, zpk2tf, zpk2sos)
9
+ from cupyx.scipy.signal._iir_filter_conversions import (
10
+ buttap, cheb1ap, cheb2ap, ellipap, buttord, ellipord, cheb1ord, cheb2ord,
11
+ _validate_gpass_gstop)
12
+
13
+
14
+ # FIXME
15
+
16
+ def besselap():
17
+ raise NotImplementedError
18
+
19
+
20
+ bessel_norms = {'fix': 'me'}
21
+
22
+
23
+ def iirfilter(N, Wn, rp=None, rs=None, btype='band', analog=False,
24
+ ftype='butter', output='ba', fs=None):
25
+ """
26
+ IIR digital and analog filter design given order and critical points.
27
+
28
+ Design an Nth-order digital or analog filter and return the filter
29
+ coefficients.
30
+
31
+ Parameters
32
+ ----------
33
+ N : int
34
+ The order of the filter.
35
+ Wn : array_like
36
+ A scalar or length-2 sequence giving the critical frequencies.
37
+
38
+ For digital filters, `Wn` are in the same units as `fs`. By default,
39
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
40
+ where 1 is the Nyquist frequency. (`Wn` is thus in
41
+ half-cycles / sample.)
42
+
43
+ For analog filters, `Wn` is an angular frequency (e.g., rad/s).
44
+
45
+ When Wn is a length-2 sequence, ``Wn[0]`` must be less than ``Wn[1]``.
46
+ rp : float, optional
47
+ For Chebyshev and elliptic filters, provides the maximum ripple
48
+ in the passband. (dB)
49
+ rs : float, optional
50
+ For Chebyshev and elliptic filters, provides the minimum attenuation
51
+ in the stop band. (dB)
52
+ btype : {'bandpass', 'lowpass', 'highpass', 'bandstop'}, optional
53
+ The type of filter. Default is 'bandpass'.
54
+ analog : bool, optional
55
+ When True, return an analog filter, otherwise a digital filter is
56
+ returned.
57
+ ftype : str, optional
58
+ The type of IIR filter to design:
59
+
60
+ - Butterworth : 'butter'
61
+ - Chebyshev I : 'cheby1'
62
+ - Chebyshev II : 'cheby2'
63
+ - Cauer/elliptic: 'ellip'
64
+ - Bessel/Thomson: 'bessel'
65
+
66
+ output : {'ba', 'zpk', 'sos'}, optional
67
+ Filter form of the output:
68
+
69
+ - second-order sections (recommended): 'sos'
70
+ - numerator/denominator (default) : 'ba'
71
+ - pole-zero : 'zpk'
72
+
73
+ In general the second-order sections ('sos') form is
74
+ recommended because inferring the coefficients for the
75
+ numerator/denominator form ('ba') suffers from numerical
76
+ instabilities. For reasons of backward compatibility the default
77
+ form is the numerator/denominator form ('ba'), where the 'b'
78
+ and the 'a' in 'ba' refer to the commonly used names of the
79
+ coefficients used.
80
+
81
+ Note: Using the second-order sections form ('sos') is sometimes
82
+ associated with additional computational costs: for
83
+ data-intense use cases it is therefore recommended to also
84
+ investigate the numerator/denominator form ('ba').
85
+
86
+ fs : float, optional
87
+ The sampling frequency of the digital system.
88
+
89
+ Returns
90
+ -------
91
+ b, a : ndarray, ndarray
92
+ Numerator (`b`) and denominator (`a`) polynomials of the IIR filter.
93
+ Only returned if ``output='ba'``.
94
+ z, p, k : ndarray, ndarray, float
95
+ Zeros, poles, and system gain of the IIR filter transfer
96
+ function. Only returned if ``output='zpk'``.
97
+ sos : ndarray
98
+ Second-order sections representation of the IIR filter.
99
+ Only returned if ``output='sos'``.
100
+
101
+ See Also
102
+ --------
103
+ butter : Filter design using order and critical points
104
+ cheby1, cheby2, ellip, bessel
105
+ buttord : Find order and critical points from passband and stopband spec
106
+ cheb1ord, cheb2ord, ellipord
107
+ iirdesign : General filter design using passband and stopband spec
108
+ scipy.signal.iirfilter
109
+
110
+ """
111
+ ftype, btype, output = [x.lower() for x in (ftype, btype, output)]
112
+
113
+ Wn = cupy.asarray(Wn)
114
+ # if cupy.any(Wn <= 0):
115
+ # raise ValueError("filter critical frequencies must be greater than 0")
116
+
117
+ if Wn.size > 1 and not Wn[0] < Wn[1]:
118
+ raise ValueError("Wn[0] must be less than Wn[1]")
119
+
120
+ if fs is not None:
121
+ if analog:
122
+ raise ValueError("fs cannot be specified for an analog filter")
123
+ Wn = 2*Wn/fs
124
+
125
+ try:
126
+ btype = band_dict[btype]
127
+ except KeyError as e:
128
+ raise ValueError(
129
+ "'%s' is an invalid bandtype for filter." % btype) from e
130
+
131
+ try:
132
+ typefunc = filter_dict[ftype][0]
133
+ except KeyError as e:
134
+ raise ValueError(
135
+ "'%s' is not a valid basic IIR filter." % ftype) from e
136
+
137
+ if output not in ['ba', 'zpk', 'sos']:
138
+ raise ValueError("'%s' is not a valid output form." % output)
139
+
140
+ if rp is not None and rp < 0:
141
+ raise ValueError("passband ripple (rp) must be positive")
142
+
143
+ if rs is not None and rs < 0:
144
+ raise ValueError("stopband attenuation (rs) must be positive")
145
+
146
+ # Get analog lowpass prototype
147
+ if typefunc == buttap:
148
+ z, p, k = typefunc(N)
149
+ elif typefunc == besselap:
150
+ z, p, k = typefunc(N, norm=bessel_norms[ftype])
151
+ elif typefunc == cheb1ap:
152
+ if rp is None:
153
+ raise ValueError("passband ripple (rp) must be provided to "
154
+ "design a Chebyshev I filter.")
155
+ z, p, k = typefunc(N, rp)
156
+ elif typefunc == cheb2ap:
157
+ if rs is None:
158
+ raise ValueError("stopband attenuation (rs) must be provided to "
159
+ "design an Chebyshev II filter.")
160
+ z, p, k = typefunc(N, rs)
161
+ elif typefunc == ellipap:
162
+ if rs is None or rp is None:
163
+ raise ValueError("Both rp and rs must be provided to design an "
164
+ "elliptic filter.")
165
+ z, p, k = typefunc(N, rp, rs)
166
+ else:
167
+ raise NotImplementedError("'%s' not implemented in iirfilter." % ftype)
168
+
169
+ # Pre-warp frequencies for digital filter design
170
+ if not analog:
171
+ if cupy.any(Wn <= 0) or cupy.any(Wn >= 1):
172
+ if fs is not None:
173
+ raise ValueError("Digital filter critical frequencies must "
174
+ f"be 0 < Wn < fs/2 (fs={fs} -> fs/2={fs/2})")
175
+ raise ValueError("Digital filter critical frequencies "
176
+ "must be 0 < Wn < 1")
177
+ fs = 2.0
178
+ warped = 2 * fs * cupy.tan(pi * Wn / fs)
179
+ else:
180
+ warped = Wn
181
+
182
+ # transform to lowpass, bandpass, highpass, or bandstop
183
+ if btype in ('lowpass', 'highpass'):
184
+ if cupy.size(Wn) != 1:
185
+ raise ValueError('Must specify a single critical frequency Wn '
186
+ 'for lowpass or highpass filter')
187
+
188
+ if btype == 'lowpass':
189
+ z, p, k = lp2lp_zpk(z, p, k, wo=warped)
190
+ elif btype == 'highpass':
191
+ z, p, k = lp2hp_zpk(z, p, k, wo=warped)
192
+ elif btype in ('bandpass', 'bandstop'):
193
+ try:
194
+ bw = warped[1] - warped[0]
195
+ wo = cupy.sqrt(warped[0] * warped[1])
196
+ except IndexError as e:
197
+ raise ValueError('Wn must specify start and stop frequencies for '
198
+ 'bandpass or bandstop filter') from e
199
+
200
+ if btype == 'bandpass':
201
+ z, p, k = lp2bp_zpk(z, p, k, wo=wo, bw=bw)
202
+ elif btype == 'bandstop':
203
+ z, p, k = lp2bs_zpk(z, p, k, wo=wo, bw=bw)
204
+ else:
205
+ raise NotImplementedError("'%s' not implemented in iirfilter." % btype)
206
+
207
+ # Find discrete equivalent if necessary
208
+ if not analog:
209
+ z, p, k = bilinear_zpk(z, p, k, fs=fs)
210
+
211
+ # Transform to proper out type (pole-zero, state-space, numer-denom)
212
+ if output == 'zpk':
213
+ return z, p, k
214
+ elif output == 'ba':
215
+ return zpk2tf(z, p, k)
216
+ elif output == 'sos':
217
+ return zpk2sos(z, p, k, analog=analog)
218
+
219
+
220
+ def butter(N, Wn, btype='low', analog=False, output='ba', fs=None):
221
+ """
222
+ Butterworth digital and analog filter design.
223
+
224
+ Design an Nth-order digital or analog Butterworth filter and return
225
+ the filter coefficients.
226
+
227
+ Parameters
228
+ ----------
229
+ N : int
230
+ The order of the filter. For 'bandpass' and 'bandstop' filters,
231
+ the resulting order of the final second-order sections ('sos')
232
+ matrix is ``2*N``, with `N` the number of biquad sections
233
+ of the desired system.
234
+ Wn : array_like
235
+ The critical frequency or frequencies. For lowpass and highpass
236
+ filters, Wn is a scalar; for bandpass and bandstop filters,
237
+ Wn is a length-2 sequence.
238
+
239
+ For a Butterworth filter, this is the point at which the gain
240
+ drops to 1/sqrt(2) that of the passband (the "-3 dB point").
241
+
242
+ For digital filters, if `fs` is not specified, `Wn` units are
243
+ normalized from 0 to 1, where 1 is the Nyquist frequency (`Wn` is
244
+ thus in half cycles / sample and defined as 2*critical frequencies
245
+ / `fs`). If `fs` is specified, `Wn` is in the same units as `fs`.
246
+
247
+ For analog filters, `Wn` is an angular frequency (e.g. rad/s).
248
+ btype : {'lowpass', 'highpass', 'bandpass', 'bandstop'}, optional
249
+ The type of filter. Default is 'lowpass'.
250
+ analog : bool, optional
251
+ When True, return an analog filter, otherwise a digital filter is
252
+ returned.
253
+ output : {'ba', 'zpk', 'sos'}, optional
254
+ Type of output: numerator/denominator ('ba'), pole-zero ('zpk'), or
255
+ second-order sections ('sos'). Default is 'ba' for backwards
256
+ compatibility, but 'sos' should be used for general-purpose filtering.
257
+ fs : float, optional
258
+ The sampling frequency of the digital system.
259
+
260
+ Returns
261
+ -------
262
+ b, a : ndarray, ndarray
263
+ Numerator (`b`) and denominator (`a`) polynomials of the IIR filter.
264
+ Only returned if ``output='ba'``.
265
+ z, p, k : ndarray, ndarray, float
266
+ Zeros, poles, and system gain of the IIR filter transfer
267
+ function. Only returned if ``output='zpk'``.
268
+ sos : ndarray
269
+ Second-order sections representation of the IIR filter.
270
+ Only returned if ``output='sos'``.
271
+
272
+ See Also
273
+ --------
274
+ buttord, buttap
275
+ iirfilter
276
+ scipy.signal.butter
277
+
278
+
279
+ Notes
280
+ -----
281
+ The Butterworth filter has maximally flat frequency response in the
282
+ passband.
283
+
284
+ If the transfer function form ``[b, a]`` is requested, numerical
285
+ problems can occur since the conversion between roots and
286
+ the polynomial coefficients is a numerically sensitive operation,
287
+ even for N >= 4. It is recommended to work with the SOS
288
+ representation.
289
+
290
+ .. warning::
291
+ Designing high-order and narrowband IIR filters in TF form can
292
+ result in unstable or incorrect filtering due to floating point
293
+ numerical precision issues. Consider inspecting output filter
294
+ characteristics `freqz` or designing the filters with second-order
295
+ sections via ``output='sos'``.
296
+ """
297
+ return iirfilter(N, Wn, btype=btype, analog=analog,
298
+ output=output, ftype='butter', fs=fs)
299
+
300
+
301
+ def cheby1(N, rp, Wn, btype='low', analog=False, output='ba', fs=None):
302
+ """
303
+ Chebyshev type I digital and analog filter design.
304
+
305
+ Design an Nth-order digital or analog Chebyshev type I filter and
306
+ return the filter coefficients.
307
+
308
+ Parameters
309
+ ----------
310
+ N : int
311
+ The order of the filter.
312
+ rp : float
313
+ The maximum ripple allowed below unity gain in the passband.
314
+ Specified in decibels, as a positive number.
315
+ Wn : array_like
316
+ A scalar or length-2 sequence giving the critical frequencies.
317
+ For Type I filters, this is the point in the transition band at which
318
+ the gain first drops below -`rp`.
319
+
320
+ For digital filters, `Wn` are in the same units as `fs`. By default,
321
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
322
+ where 1 is the Nyquist frequency. (`Wn` is thus in
323
+ half-cycles / sample.)
324
+
325
+ For analog filters, `Wn` is an angular frequency (e.g., rad/s).
326
+ btype : {'lowpass', 'highpass', 'bandpass', 'bandstop'}, optional
327
+ The type of filter. Default is 'lowpass'.
328
+ analog : bool, optional
329
+ When True, return an analog filter, otherwise a digital filter is
330
+ returned.
331
+ output : {'ba', 'zpk', 'sos'}, optional
332
+ Type of output: numerator/denominator ('ba'), pole-zero ('zpk'), or
333
+ second-order sections ('sos'). Default is 'ba' for backwards
334
+ compatibility, but 'sos' should be used for general-purpose filtering.
335
+ fs : float, optional
336
+ The sampling frequency of the digital system.
337
+
338
+ Returns
339
+ -------
340
+ b, a : ndarray, ndarray
341
+ Numerator (`b`) and denominator (`a`) polynomials of the IIR filter.
342
+ Only returned if ``output='ba'``.
343
+ z, p, k : ndarray, ndarray, float
344
+ Zeros, poles, and system gain of the IIR filter transfer
345
+ function. Only returned if ``output='zpk'``.
346
+ sos : ndarray
347
+ Second-order sections representation of the IIR filter.
348
+ Only returned if ``output='sos'``.
349
+
350
+ See Also
351
+ --------
352
+ cheb1ord, cheb1ap
353
+ iirfilter
354
+ scipy.signal.cheby1
355
+
356
+ Notes
357
+ -----
358
+ The Chebyshev type I filter maximizes the rate of cutoff between the
359
+ frequency response's passband and stopband, at the expense of ripple in
360
+ the passband and increased ringing in the step response.
361
+
362
+ Type I filters roll off faster than Type II (`cheby2`), but Type II
363
+ filters do not have any ripple in the passband.
364
+
365
+ The equiripple passband has N maxima or minima (for example, a
366
+ 5th-order filter has 3 maxima and 2 minima). Consequently, the DC gain is
367
+ unity for odd-order filters, or -rp dB for even-order filters.
368
+ """
369
+ return iirfilter(N, Wn, rp=rp, btype=btype, analog=analog,
370
+ output=output, ftype='cheby1', fs=fs)
371
+
372
+
373
+ def cheby2(N, rs, Wn, btype='low', analog=False, output='ba', fs=None):
374
+ """
375
+ Chebyshev type II digital and analog filter design.
376
+
377
+ Design an Nth-order digital or analog Chebyshev type II filter and
378
+ return the filter coefficients.
379
+
380
+ Parameters
381
+ ----------
382
+ N : int
383
+ The order of the filter.
384
+ rs : float
385
+ The minimum attenuation required in the stop band.
386
+ Specified in decibels, as a positive number.
387
+ Wn : array_like
388
+ A scalar or length-2 sequence giving the critical frequencies.
389
+ For Type II filters, this is the point in the transition band at which
390
+ the gain first reaches -`rs`.
391
+
392
+ For digital filters, `Wn` are in the same units as `fs`. By default,
393
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
394
+ where 1 is the Nyquist frequency. (`Wn` is thus in
395
+ half-cycles / sample.)
396
+
397
+ For analog filters, `Wn` is an angular frequency (e.g., rad/s).
398
+ btype : {'lowpass', 'highpass', 'bandpass', 'bandstop'}, optional
399
+ The type of filter. Default is 'lowpass'.
400
+ analog : bool, optional
401
+ When True, return an analog filter, otherwise a digital filter is
402
+ returned.
403
+ output : {'ba', 'zpk', 'sos'}, optional
404
+ Type of output: numerator/denominator ('ba'), pole-zero ('zpk'), or
405
+ second-order sections ('sos'). Default is 'ba' for backwards
406
+ compatibility, but 'sos' should be used for general-purpose filtering.
407
+ fs : float, optional
408
+ The sampling frequency of the digital system.
409
+
410
+ Returns
411
+ -------
412
+ b, a : ndarray, ndarray
413
+ Numerator (`b`) and denominator (`a`) polynomials of the IIR filter.
414
+ Only returned if ``output='ba'``.
415
+ z, p, k : ndarray, ndarray, float
416
+ Zeros, poles, and system gain of the IIR filter transfer
417
+ function. Only returned if ``output='zpk'``.
418
+ sos : ndarray
419
+ Second-order sections representation of the IIR filter.
420
+ Only returned if ``output='sos'``.
421
+
422
+ See Also
423
+ --------
424
+ cheb2ord, cheb2ap
425
+ iirfilter
426
+ scipy.signal.cheby2
427
+
428
+ Notes
429
+ -----
430
+ The Chebyshev type II filter maximizes the rate of cutoff between the
431
+ frequency response's passband and stopband, at the expense of ripple in
432
+ the stopband and increased ringing in the step response.
433
+
434
+ Type II filters do not roll off as fast as Type I (`cheby1`).
435
+ """
436
+ return iirfilter(N, Wn, rs=rs, btype=btype, analog=analog,
437
+ output=output, ftype='cheby2', fs=fs)
438
+
439
+
440
+ def ellip(N, rp, rs, Wn, btype='low', analog=False, output='ba', fs=None):
441
+ """
442
+ Elliptic (Cauer) digital and analog filter design.
443
+
444
+ Design an Nth-order digital or analog elliptic filter and return
445
+ the filter coefficients.
446
+
447
+ Parameters
448
+ ----------
449
+ N : int
450
+ The order of the filter.
451
+ rp : float
452
+ The maximum ripple allowed below unity gain in the passband.
453
+ Specified in decibels, as a positive number.
454
+ rs : float
455
+ The minimum attenuation required in the stop band.
456
+ Specified in decibels, as a positive number.
457
+ Wn : array_like
458
+ A scalar or length-2 sequence giving the critical frequencies.
459
+ For elliptic filters, this is the point in the transition band at
460
+ which the gain first drops below -`rp`.
461
+
462
+ For digital filters, `Wn` are in the same units as `fs`. By default,
463
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
464
+ where 1 is the Nyquist frequency. (`Wn` is thus in
465
+ half-cycles / sample.)
466
+
467
+ For analog filters, `Wn` is an angular frequency (e.g., rad/s).
468
+ btype : {'lowpass', 'highpass', 'bandpass', 'bandstop'}, optional
469
+ The type of filter. Default is 'lowpass'.
470
+ analog : bool, optional
471
+ When True, return an analog filter, otherwise a digital filter is
472
+ returned.
473
+ output : {'ba', 'zpk', 'sos'}, optional
474
+ Type of output: numerator/denominator ('ba'), pole-zero ('zpk'), or
475
+ second-order sections ('sos'). Default is 'ba' for backwards
476
+ compatibility, but 'sos' should be used for general-purpose filtering.
477
+ fs : float, optional
478
+ The sampling frequency of the digital system.
479
+
480
+ Returns
481
+ -------
482
+ b, a : ndarray, ndarray
483
+ Numerator (`b`) and denominator (`a`) polynomials of the IIR filter.
484
+ Only returned if ``output='ba'``.
485
+ z, p, k : ndarray, ndarray, float
486
+ Zeros, poles, and system gain of the IIR filter transfer
487
+ function. Only returned if ``output='zpk'``.
488
+ sos : ndarray
489
+ Second-order sections representation of the IIR filter.
490
+ Only returned if ``output='sos'``.
491
+
492
+ See Also
493
+ --------
494
+ ellipord, ellipap
495
+ iirfilter
496
+ scipy.signal.ellip
497
+
498
+ Notes
499
+ -----
500
+ Also known as Cauer or Zolotarev filters, the elliptical filter maximizes
501
+ the rate of transition between the frequency response's passband and
502
+ stopband, at the expense of ripple in both, and increased ringing in the
503
+ step response.
504
+
505
+ As `rp` approaches 0, the elliptical filter becomes a Chebyshev
506
+ type II filter (`cheby2`). As `rs` approaches 0, it becomes a Chebyshev
507
+ type I filter (`cheby1`). As both approach 0, it becomes a Butterworth
508
+ filter (`butter`).
509
+
510
+ The equiripple passband has N maxima or minima (for example, a
511
+ 5th-order filter has 3 maxima and 2 minima). Consequently, the DC gain is
512
+ unity for odd-order filters, or -rp dB for even-order filters.
513
+ """
514
+ return iirfilter(N, Wn, rs=rs, rp=rp, btype=btype, analog=analog,
515
+ output=output, ftype='elliptic', fs=fs)
516
+
517
+
518
+ def iirdesign(wp, ws, gpass, gstop, analog=False, ftype='ellip', output='ba',
519
+ fs=None):
520
+ """Complete IIR digital and analog filter design.
521
+
522
+ Given passband and stopband frequencies and gains, construct an analog or
523
+ digital IIR filter of minimum order for a given basic type. Return the
524
+ output in numerator, denominator ('ba'), pole-zero ('zpk') or second order
525
+ sections ('sos') form.
526
+
527
+ Parameters
528
+ ----------
529
+ wp, ws : float or array like, shape (2,)
530
+ Passband and stopband edge frequencies. Possible values are scalars
531
+ (for lowpass and highpass filters) or ranges (for bandpass and bandstop
532
+ filters).
533
+ For digital filters, these are in the same units as `fs`. By default,
534
+ `fs` is 2 half-cycles/sample, so these are normalized from 0 to 1,
535
+ where 1 is the Nyquist frequency. For example:
536
+
537
+ - Lowpass: wp = 0.2, ws = 0.3
538
+ - Highpass: wp = 0.3, ws = 0.2
539
+ - Bandpass: wp = [0.2, 0.5], ws = [0.1, 0.6]
540
+ - Bandstop: wp = [0.1, 0.6], ws = [0.2, 0.5]
541
+
542
+ For analog filters, `wp` and `ws` are angular frequencies
543
+ (e.g., rad/s). Note, that for bandpass and bandstop filters passband
544
+ must lie strictly inside stopband or vice versa.
545
+ gpass : float
546
+ The maximum loss in the passband (dB).
547
+ gstop : float
548
+ The minimum attenuation in the stopband (dB).
549
+ analog : bool, optional
550
+ When True, return an analog filter, otherwise a digital filter is
551
+ returned.
552
+ ftype : str, optional
553
+ The type of IIR filter to design:
554
+
555
+ - Butterworth : 'butter'
556
+ - Chebyshev I : 'cheby1'
557
+ - Chebyshev II : 'cheby2'
558
+ - Cauer/elliptic: 'ellip'
559
+
560
+ output : {'ba', 'zpk', 'sos'}, optional
561
+ Filter form of the output:
562
+
563
+ - second-order sections (recommended): 'sos'
564
+ - numerator/denominator (default) : 'ba'
565
+ - pole-zero : 'zpk'
566
+
567
+ In general the second-order sections ('sos') form is
568
+ recommended because inferring the coefficients for the
569
+ numerator/denominator form ('ba') suffers from numerical
570
+ instabilities. For reasons of backward compatibility the default
571
+ form is the numerator/denominator form ('ba'), where the 'b'
572
+ and the 'a' in 'ba' refer to the commonly used names of the
573
+ coefficients used.
574
+
575
+ Note: Using the second-order sections form ('sos') is sometimes
576
+ associated with additional computational costs: for
577
+ data-intense use cases it is therefore recommended to also
578
+ investigate the numerator/denominator form ('ba').
579
+
580
+ fs : float, optional
581
+ The sampling frequency of the digital system.
582
+
583
+ .. versionadded:: 1.2.0
584
+
585
+ Returns
586
+ -------
587
+ b, a : ndarray, ndarray
588
+ Numerator (`b`) and denominator (`a`) polynomials of the IIR filter.
589
+ Only returned if ``output='ba'``.
590
+ z, p, k : ndarray, ndarray, float
591
+ Zeros, poles, and system gain of the IIR filter transfer
592
+ function. Only returned if ``output='zpk'``.
593
+ sos : ndarray
594
+ Second-order sections representation of the IIR filter.
595
+ Only returned if ``output='sos'``.
596
+
597
+ See Also
598
+ --------
599
+ scipy.signal.iirdesign
600
+ butter : Filter design using order and critical points
601
+ cheby1, cheby2, ellip, bessel
602
+ buttord : Find order and critical points from passband and stopband spec
603
+ cheb1ord, cheb2ord, ellipord
604
+ iirfilter : General filter design using order and critical frequencies
605
+ """
606
+ try:
607
+ ordfunc = filter_dict[ftype][1]
608
+ except KeyError as e:
609
+ raise ValueError("Invalid IIR filter type: %s" % ftype) from e
610
+ except IndexError as e:
611
+ raise ValueError(("%s does not have order selection. Use "
612
+ "iirfilter function.") % ftype) from e
613
+
614
+ _validate_gpass_gstop(gpass, gstop)
615
+
616
+ wp = cupy.atleast_1d(wp)
617
+ ws = cupy.atleast_1d(ws)
618
+
619
+ if wp.shape[0] != ws.shape[0] or wp.shape not in [(1,), (2,)]:
620
+ raise ValueError("wp and ws must have one or two elements each, and"
621
+ "the same shape, got %s and %s"
622
+ % (wp.shape, ws.shape))
623
+
624
+ if any(wp <= 0) or any(ws <= 0):
625
+ raise ValueError("Values for wp, ws must be greater than 0")
626
+
627
+ if not analog:
628
+ if fs is None:
629
+ if any(wp >= 1) or any(ws >= 1):
630
+ raise ValueError("Values for wp, ws must be less than 1")
631
+ elif any(wp >= fs/2) or any(ws >= fs/2):
632
+ raise ValueError("Values for wp, ws must be less than fs/2"
633
+ " (fs={} -> fs/2={})".format(fs, fs/2))
634
+
635
+ if wp.shape[0] == 2:
636
+ if not ((ws[0] < wp[0] and wp[1] < ws[1]) or
637
+ (wp[0] < ws[0] and ws[1] < wp[1])):
638
+ raise ValueError("Passband must lie strictly inside stopband"
639
+ " or vice versa")
640
+
641
+ band_type = 2 * (len(wp) - 1)
642
+ band_type += 1
643
+ if wp[0] >= ws[0]:
644
+ band_type += 1
645
+
646
+ btype = {1: 'lowpass', 2: 'highpass',
647
+ 3: 'bandstop', 4: 'bandpass'}[band_type]
648
+
649
+ N, Wn = ordfunc(wp, ws, gpass, gstop, analog=analog, fs=fs)
650
+ return iirfilter(N, Wn, rp=gpass, rs=gstop, analog=analog, btype=btype,
651
+ ftype=ftype, output=output, fs=fs)
652
+
653
+
654
+ def iircomb(w0, Q, ftype='notch', fs=2.0, *, pass_zero=False):
655
+ """
656
+ Design IIR notching or peaking digital comb filter.
657
+
658
+ A notching comb filter consists of regularly-spaced band-stop filters with
659
+ a narrow bandwidth (high quality factor). Each rejects a narrow frequency
660
+ band and leaves the rest of the spectrum little changed.
661
+
662
+ A peaking comb filter consists of regularly-spaced band-pass filters with
663
+ a narrow bandwidth (high quality factor). Each rejects components outside
664
+ a narrow frequency band.
665
+
666
+ Parameters
667
+ ----------
668
+ w0 : float
669
+ The fundamental frequency of the comb filter (the spacing between its
670
+ peaks). This must evenly divide the sampling frequency. If `fs` is
671
+ specified, this is in the same units as `fs`. By default, it is
672
+ a normalized scalar that must satisfy ``0 < w0 < 1``, with
673
+ ``w0 = 1`` corresponding to half of the sampling frequency.
674
+ Q : float
675
+ Quality factor. Dimensionless parameter that characterizes
676
+ notch filter -3 dB bandwidth ``bw`` relative to its center
677
+ frequency, ``Q = w0/bw``.
678
+ ftype : {'notch', 'peak'}
679
+ The type of comb filter generated by the function. If 'notch', then
680
+ the Q factor applies to the notches. If 'peak', then the Q factor
681
+ applies to the peaks. Default is 'notch'.
682
+ fs : float, optional
683
+ The sampling frequency of the signal. Default is 2.0.
684
+ pass_zero : bool, optional
685
+ If False (default), the notches (nulls) of the filter are centered on
686
+ frequencies [0, w0, 2*w0, ...], and the peaks are centered on the
687
+ midpoints [w0/2, 3*w0/2, 5*w0/2, ...]. If True, the peaks are centered
688
+ on [0, w0, 2*w0, ...] (passing zero frequency) and vice versa.
689
+
690
+ Returns
691
+ -------
692
+ b, a : ndarray, ndarray
693
+ Numerator (``b``) and denominator (``a``) polynomials
694
+ of the IIR filter.
695
+
696
+ Raises
697
+ ------
698
+ ValueError
699
+ If `w0` is less than or equal to 0 or greater than or equal to
700
+ ``fs/2``, if `fs` is not divisible by `w0`, if `ftype`
701
+ is not 'notch' or 'peak'
702
+
703
+ See Also
704
+ --------
705
+ scipy.signal.iircomb
706
+ iirnotch
707
+ iirpeak
708
+
709
+ Notes
710
+ -----
711
+ The TF implementation of the
712
+ comb filter is numerically stable even at higher orders due to the
713
+ use of a single repeated pole, which won't suffer from precision loss.
714
+
715
+ References
716
+ ----------
717
+ Sophocles J. Orfanidis, "Introduction To Signal Processing",
718
+ Prentice-Hall, 1996, ch. 11, "Digital Filter Design"
719
+ """
720
+
721
+ # Convert w0, Q, and fs to float
722
+ w0 = float(w0)
723
+ Q = float(Q)
724
+ fs = float(fs)
725
+
726
+ # Check for invalid cutoff frequency or filter type
727
+ ftype = ftype.lower()
728
+ if not 0 < w0 < fs / 2:
729
+ raise ValueError("w0 must be between 0 and {}"
730
+ " (nyquist), but given {}.".format(fs / 2, w0))
731
+ if ftype not in ('notch', 'peak'):
732
+ raise ValueError('ftype must be either notch or peak.')
733
+
734
+ # Compute the order of the filter
735
+ N = round(fs / w0)
736
+
737
+ # Check for cutoff frequency divisibility
738
+ if abs(w0 - fs/N)/fs > 1e-14:
739
+ raise ValueError('fs must be divisible by w0.')
740
+
741
+ # Compute frequency in radians and filter bandwidth
742
+ # Eq. 11.3.1 (p. 574) from reference [1]
743
+ w0 = (2 * pi * w0) / fs
744
+ w_delta = w0 / Q
745
+
746
+ # Define base gain values depending on notch or peak filter
747
+ # Compute -3dB attenuation
748
+ # Eqs. 11.4.1 and 11.4.2 (p. 582) from reference [1]
749
+ if ftype == 'notch':
750
+ G0, G = 1, 0
751
+ elif ftype == 'peak':
752
+ G0, G = 0, 1
753
+ GB = 1 / math.sqrt(2)
754
+
755
+ # Compute beta
756
+ # Eq. 11.5.3 (p. 591) from reference [1]
757
+ beta = math.sqrt((GB**2 - G0**2) / (G**2 - GB**2)) * \
758
+ math.tan(N * w_delta / 4)
759
+
760
+ # Compute filter coefficients
761
+ # Eq 11.5.1 (p. 590) variables a, b, c from reference [1]
762
+ ax = (1 - beta) / (1 + beta)
763
+ bx = (G0 + G * beta) / (1 + beta)
764
+ cx = (G0 - G * beta) / (1 + beta)
765
+
766
+ # Last coefficients are negative to get peaking comb that passes zero or
767
+ # notching comb that doesn't.
768
+ negative_coef = ((ftype == 'peak' and pass_zero) or
769
+ (ftype == 'notch' and not pass_zero))
770
+
771
+ # Compute numerator coefficients
772
+ # Eq 11.5.1 (p. 590) or Eq 11.5.4 (p. 591) from reference [1]
773
+ # b - cz^-N or b + cz^-N
774
+ b = cupy.zeros(N + 1)
775
+ b[0] = bx
776
+ if negative_coef:
777
+ b[-1] = -cx
778
+ else:
779
+ b[-1] = +cx
780
+
781
+ # Compute denominator coefficients
782
+ # Eq 11.5.1 (p. 590) or Eq 11.5.4 (p. 591) from reference [1]
783
+ # 1 - az^-N or 1 + az^-N
784
+ a = cupy.zeros(N + 1)
785
+ a[0] = 1
786
+ if negative_coef:
787
+ a[-1] = -ax
788
+ else:
789
+ a[-1] = +ax
790
+
791
+ return b, a
792
+
793
+
794
+ def iirnotch(w0, Q, fs=2.0):
795
+ """
796
+ Design second-order IIR notch digital filter.
797
+
798
+ A notch filter is a band-stop filter with a narrow bandwidth
799
+ (high quality factor). It rejects a narrow frequency band and
800
+ leaves the rest of the spectrum little changed.
801
+
802
+ Parameters
803
+ ----------
804
+ w0 : float
805
+ Frequency to remove from a signal. If `fs` is specified, this is in
806
+ the same units as `fs`. By default, it is a normalized scalar that must
807
+ satisfy ``0 < w0 < 1``, with ``w0 = 1`` corresponding to half of the
808
+ sampling frequency.
809
+ Q : float
810
+ Quality factor. Dimensionless parameter that characterizes
811
+ notch filter -3 dB bandwidth ``bw`` relative to its center
812
+ frequency, ``Q = w0/bw``.
813
+ fs : float, optional
814
+ The sampling frequency of the digital system.
815
+
816
+ Returns
817
+ -------
818
+ b, a : ndarray, ndarray
819
+ Numerator (``b``) and denominator (``a``) polynomials
820
+ of the IIR filter.
821
+
822
+ See Also
823
+ --------
824
+ scipy.signal.iirnotch
825
+
826
+ References
827
+ ----------
828
+ Sophocles J. Orfanidis, "Introduction To Signal Processing",
829
+ Prentice-Hall, 1996
830
+ """
831
+
832
+ return _design_notch_peak_filter(w0, Q, "notch", fs)
833
+
834
+
835
+ def iirpeak(w0, Q, fs=2.0):
836
+ """
837
+ Design second-order IIR peak (resonant) digital filter.
838
+
839
+ A peak filter is a band-pass filter with a narrow bandwidth
840
+ (high quality factor). It rejects components outside a narrow
841
+ frequency band.
842
+
843
+ Parameters
844
+ ----------
845
+ w0 : float
846
+ Frequency to be retained in a signal. If `fs` is specified, this is in
847
+ the same units as `fs`. By default, it is a normalized scalar that must
848
+ satisfy ``0 < w0 < 1``, with ``w0 = 1`` corresponding to half of the
849
+ sampling frequency.
850
+ Q : float
851
+ Quality factor. Dimensionless parameter that characterizes
852
+ peak filter -3 dB bandwidth ``bw`` relative to its center
853
+ frequency, ``Q = w0/bw``.
854
+ fs : float, optional
855
+ The sampling frequency of the digital system.
856
+
857
+
858
+ Returns
859
+ -------
860
+ b, a : ndarray, ndarray
861
+ Numerator (``b``) and denominator (``a``) polynomials
862
+ of the IIR filter.
863
+
864
+ See Also
865
+ --------
866
+ scpy.signal.iirpeak
867
+
868
+ References
869
+ ----------
870
+ Sophocles J. Orfanidis, "Introduction To Signal Processing",
871
+ Prentice-Hall, 1996
872
+ """
873
+
874
+ return _design_notch_peak_filter(w0, Q, "peak", fs)
875
+
876
+
877
+ def _design_notch_peak_filter(w0, Q, ftype, fs=2.0):
878
+ """
879
+ Design notch or peak digital filter.
880
+
881
+ Parameters
882
+ ----------
883
+ w0 : float
884
+ Normalized frequency to remove from a signal. If `fs` is specified,
885
+ this is in the same units as `fs`. By default, it is a normalized
886
+ scalar that must satisfy ``0 < w0 < 1``, with ``w0 = 1``
887
+ corresponding to half of the sampling frequency.
888
+ Q : float
889
+ Quality factor. Dimensionless parameter that characterizes
890
+ notch filter -3 dB bandwidth ``bw`` relative to its center
891
+ frequency, ``Q = w0/bw``.
892
+ ftype : str
893
+ The type of IIR filter to design:
894
+
895
+ - notch filter : ``notch``
896
+ - peak filter : ``peak``
897
+ fs : float, optional
898
+ The sampling frequency of the digital system.
899
+
900
+ Returns
901
+ -------
902
+ b, a : ndarray, ndarray
903
+ Numerator (``b``) and denominator (``a``) polynomials
904
+ of the IIR filter.
905
+ """
906
+
907
+ # Guarantee that the inputs are floats
908
+ w0 = float(w0)
909
+ Q = float(Q)
910
+ w0 = 2 * w0 / fs
911
+
912
+ # Checks if w0 is within the range
913
+ if w0 > 1.0 or w0 < 0.0:
914
+ raise ValueError("w0 should be such that 0 < w0 < 1")
915
+
916
+ # Get bandwidth
917
+ bw = w0 / Q
918
+
919
+ # Normalize inputs
920
+ bw = bw * pi
921
+ w0 = w0 * pi
922
+
923
+ # Compute -3dB attenuation
924
+ gb = 1 / math.sqrt(2)
925
+
926
+ if ftype == "notch":
927
+ # Compute beta: formula 11.3.4 (p.575) from reference [1]
928
+ beta = (math.sqrt(1.0 - gb**2.0) / gb) * math.tan(bw / 2.0)
929
+ elif ftype == "peak":
930
+ # Compute beta: formula 11.3.19 (p.579) from reference [1]
931
+ beta = (gb / math.sqrt(1.0 - gb**2.0)) * math.tan(bw / 2.0)
932
+ else:
933
+ raise ValueError("Unknown ftype.")
934
+
935
+ # Compute gain: formula 11.3.6 (p.575) from reference [1]
936
+ gain = 1.0 / (1.0 + beta)
937
+
938
+ # Compute numerator b and denominator a
939
+ # formulas 11.3.7 (p.575) and 11.3.21 (p.579)
940
+ # from reference [1]
941
+ if ftype == "notch":
942
+ b = [gain * x for x in (1.0, -2.0 * math.cos(w0), 1.0)]
943
+ else:
944
+ b = [(1.0 - gain) * x for x in (1.0, 0.0, -1.0)]
945
+
946
+ a = [1.0, -2.0 * gain * math.cos(w0), 2.0 * gain - 1.0]
947
+
948
+ a = cupy.asarray(a)
949
+ b = cupy.asarray(b)
950
+
951
+ return b, a
952
+
953
+
954
+ filter_dict = {'butter': [buttap, buttord],
955
+ 'butterworth': [buttap, buttord],
956
+
957
+ 'cauer': [ellipap, ellipord],
958
+ 'elliptic': [ellipap, ellipord],
959
+ 'ellip': [ellipap, ellipord],
960
+
961
+ 'bessel': [besselap],
962
+ 'bessel_phase': [besselap],
963
+ 'bessel_delay': [besselap],
964
+ 'bessel_mag': [besselap],
965
+
966
+ 'cheby1': [cheb1ap, cheb1ord],
967
+ 'chebyshev1': [cheb1ap, cheb1ord],
968
+ 'chebyshevi': [cheb1ap, cheb1ord],
969
+
970
+ 'cheby2': [cheb2ap, cheb2ord],
971
+ 'chebyshev2': [cheb2ap, cheb2ord],
972
+ 'chebyshevii': [cheb2ap, cheb2ord],
973
+ }
974
+
975
+ band_dict = {'band': 'bandpass',
976
+ 'bandpass': 'bandpass',
977
+ 'pass': 'bandpass',
978
+ 'bp': 'bandpass',
979
+
980
+ 'bs': 'bandstop',
981
+ 'bandstop': 'bandstop',
982
+ 'bands': 'bandstop',
983
+ 'stop': 'bandstop',
984
+
985
+ 'l': 'lowpass',
986
+ 'low': 'lowpass',
987
+ 'lowpass': 'lowpass',
988
+ 'lp': 'lowpass',
989
+
990
+ 'high': 'highpass',
991
+ 'highpass': 'highpass',
992
+ 'h': 'highpass',
993
+ 'hp': 'highpass',
994
+ }
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_iir_utils.py ADDED
@@ -0,0 +1,737 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ from itertools import product
3
+
4
+ import cupy
5
+ from cupy._core.internal import _normalize_axis_index
6
+ from cupy._core._scalar import get_typename
7
+ from cupy_backends.cuda.api import runtime
8
+ from cupyx.scipy.signal._arraytools import axis_slice
9
+
10
+
11
+ def _get_typename(dtype):
12
+ typename = get_typename(dtype)
13
+ if cupy.dtype(dtype).kind == 'c':
14
+ typename = 'thrust::' + typename
15
+ elif typename == 'float16':
16
+ if runtime.is_hip:
17
+ # 'half' in name_expressions weirdly raises
18
+ # HIPRTC_ERROR_NAME_EXPRESSION_NOT_VALID in getLoweredName() on
19
+ # ROCm
20
+ typename = '__half'
21
+ else:
22
+ typename = 'half'
23
+ return typename
24
+
25
+
26
+ FLOAT_TYPES = [cupy.float16, cupy.float32, cupy.float64]
27
+ INT_TYPES = [cupy.int8, cupy.int16, cupy.int32, cupy.int64]
28
+ COMPLEX_TYPES = [cupy.complex64, cupy.complex128]
29
+ UNSIGNED_TYPES = [cupy.uint8, cupy.uint16, cupy.uint32, cupy.uint64]
30
+ TYPES = FLOAT_TYPES + INT_TYPES + UNSIGNED_TYPES + COMPLEX_TYPES # type: ignore # NOQA
31
+ TYPE_PAIRS = [(x, y) for x, y in product(TYPES, TYPES)
32
+ if cupy.promote_types(x, y) is cupy.dtype(x)]
33
+
34
+ TYPE_NAMES = [_get_typename(t) for t in TYPES]
35
+ TYPE_PAIR_NAMES = [(_get_typename(x), _get_typename(y)) for x, y in TYPE_PAIRS]
36
+
37
+
38
+ IIR_KERNEL = r"""
39
+ #include <cupy/math_constants.h>
40
+ #include <cupy/carray.cuh>
41
+ #include <cupy/complex.cuh>
42
+
43
+ template<typename U, typename T>
44
+ __global__ void compute_correction_factors(
45
+ const int m, const int k, const T* b, U* out) {
46
+ int idx = blockDim.x * blockIdx.x + threadIdx.x;
47
+ if(idx >= k) {
48
+ return;
49
+ }
50
+
51
+ U* out_start = out + idx * (k + m);
52
+ U* out_off = out_start + k;
53
+
54
+ for(int i = 0; i < m; i++) {
55
+ U acc = 0.0;
56
+ for(int j = 0; j < k; j++) {
57
+ acc += ((U) b[j]) * out_off[i - j - 1];
58
+
59
+ }
60
+ out_off[i] = acc;
61
+ }
62
+ }
63
+
64
+ template<typename T>
65
+ __global__ void first_pass_iir(
66
+ const int m, const int k, const int n, const int n_blocks,
67
+ const int carries_stride, const T* factors, T* out,
68
+ T* carries) {
69
+ int orig_idx = blockDim.x * (blockIdx.x % n_blocks) + threadIdx.x;
70
+
71
+ int num_row = blockIdx.x / n_blocks;
72
+ int idx = 2 * orig_idx + 1;
73
+
74
+ if(idx >= n) {
75
+ return;
76
+ }
77
+
78
+ int group_num = idx / m;
79
+ int group_pos = idx % m;
80
+
81
+ T* out_off = out + num_row * n;
82
+ T* carries_off = carries + num_row * carries_stride;
83
+
84
+ T* group_start = out_off + m * group_num;
85
+ T* group_carries = carries_off + k * group_num;
86
+
87
+ int pos = group_pos;
88
+ int up_bound = pos;
89
+ int low_bound = pos;
90
+ int rel_pos;
91
+
92
+ for(int level = 1, iter = 1; level < m; level *=2, iter++) {
93
+ int sz = min(pow(2.0f, ((float) iter)), ((float) m));
94
+
95
+ if(level > 1) {
96
+ int factor = ceil(pos / ((float) sz));
97
+ up_bound = sz * factor - 1;
98
+ low_bound = up_bound - level + 1;
99
+ }
100
+
101
+ if(level == 1) {
102
+ pos = low_bound;
103
+ }
104
+
105
+ if(pos < low_bound) {
106
+ pos += level / 2;
107
+ }
108
+
109
+ if(pos + m * group_num >= n) {
110
+ break;
111
+ }
112
+
113
+ rel_pos = pos % level;
114
+ T carry = 0.0;
115
+ for(int i = 1; i <= min(k, level); i++) {
116
+ T k_value = group_start[low_bound - i];
117
+ const T* k_factors = factors + (m + k) * (i - 1) + k;
118
+ T factor = k_factors[rel_pos];
119
+ carry += k_value * factor;
120
+ }
121
+
122
+ group_start[pos] += carry;
123
+ __syncthreads();
124
+ }
125
+
126
+ if(pos >= m - k) {
127
+ if(carries != NULL) {
128
+ group_carries[pos - (m - k)] = group_start[pos];
129
+ }
130
+ }
131
+
132
+ }
133
+
134
+ template<typename T>
135
+ __global__ void correct_carries(
136
+ const int m, const int k, const int n_blocks, const int carries_stride,
137
+ const int offset, const T* factors, T* carries) {
138
+
139
+ int idx = threadIdx.x;
140
+ int pos = idx + (m - k);
141
+ T* row_carries = carries + carries_stride * blockIdx.x;
142
+
143
+ for(int i = offset; i < n_blocks; i++) {
144
+ T* this_carries = row_carries + k * (i + (1 - offset));
145
+ T* prev_carries = row_carries + k * (i - offset);
146
+
147
+ T carry = 0.0;
148
+ for(int j = 1; j <= k; j++) {
149
+ const T* k_factors = factors + (m + k) * (j - 1) + k;
150
+ T factor = k_factors[pos];
151
+ T k_value = prev_carries[k - j];
152
+ carry += factor * k_value;
153
+ }
154
+
155
+ this_carries[idx] += carry;
156
+ __syncthreads();
157
+ }
158
+ }
159
+
160
+ template<typename T>
161
+ __global__ void second_pass_iir(
162
+ const int m, const int k, const int n, const int carries_stride,
163
+ const int n_blocks, const int offset, const T* factors,
164
+ T* carries, T* out) {
165
+
166
+ int idx = blockDim.x * (blockIdx.x % n_blocks) + threadIdx.x;
167
+ idx += offset * m;
168
+
169
+ int row_num = blockIdx.x / n_blocks;
170
+ int n_group = idx / m;
171
+ int pos = idx % m;
172
+
173
+ if(idx >= n) {
174
+ return;
175
+ }
176
+
177
+ T* out_off = out + row_num * n;
178
+ T* carries_off = carries + row_num * carries_stride;
179
+ const T* prev_carries = carries_off + (n_group - offset) * k;
180
+
181
+ T carry = 0.0;
182
+ for(int i = 1; i <= k; i++) {
183
+ const T* k_factors = factors + (m + k) * (i - 1) + k;
184
+ T factor = k_factors[pos];
185
+ T k_value = prev_carries[k - i];
186
+ carry += factor * k_value;
187
+ }
188
+
189
+ out_off[idx] += carry;
190
+ }
191
+ """
192
+
193
+ IIR_SOS_KERNEL = r"""
194
+ #include <cupy/math_constants.h>
195
+ #include <cupy/carray.cuh>
196
+ #include <cupy/complex.cuh>
197
+
198
+ template<typename T>
199
+ __global__ void pick_carries(
200
+ const int m, const int n, const int carries_stride, const int n_blocks,
201
+ const int offset, T* x, T* carries) {
202
+
203
+ int idx = m * (blockIdx.x % n_blocks) + threadIdx.x + m - 2;
204
+ int pos = threadIdx.x;
205
+ int row_num = blockIdx.x / n_blocks;
206
+ int n_group = idx / m;
207
+
208
+ T* x_off = x + row_num * n;
209
+ T* carries_off = carries + row_num * carries_stride;
210
+ T* group_carries = carries_off + (n_group + (1 - offset)) * 2;
211
+
212
+ if(idx >= n) {
213
+ return;
214
+ }
215
+
216
+ group_carries[pos] = x_off[idx];
217
+ }
218
+
219
+ template<typename U, typename T>
220
+ __global__ void compute_correction_factors_sos(
221
+ const int m, const T* f_const, U* all_out) {
222
+
223
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> bc_d[2];
224
+ T* b_c = reinterpret_cast<T*>(bc_d);
225
+
226
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> off_d[4];
227
+ U* off_cache = reinterpret_cast<U*>(off_d);
228
+
229
+ int idx = threadIdx.x;
230
+ int num_section = blockIdx.x;
231
+
232
+ const int n_const = 6;
233
+ const int a_off = 3;
234
+ const int k = 2;
235
+ const int off_idx = 1;
236
+
237
+ U* out = all_out + num_section * k * m;
238
+ U* out_start = out + idx * m;
239
+ const T* b = f_const + num_section * n_const + a_off + 1;
240
+
241
+ b_c[idx] = b[idx];
242
+ __syncthreads();
243
+
244
+ U* this_cache = off_cache + k * idx;
245
+ this_cache[off_idx - idx] = 1;
246
+ this_cache[idx] = 0;
247
+
248
+ for(int i = 0; i < m; i++) {
249
+ U acc = 0.0;
250
+ for(int j = 0; j < k; j++) {
251
+ acc += -((U) b_c[j]) * this_cache[off_idx - j];
252
+
253
+ }
254
+ this_cache[0] = this_cache[1];
255
+ this_cache[1] = acc;
256
+ out_start[i] = acc;
257
+ }
258
+ }
259
+
260
+
261
+ template<typename T>
262
+ __global__ void first_pass_iir_sos(
263
+ const int m, const int n, const int n_blocks,
264
+ const T* factors, T* out, T* carries) {
265
+
266
+ extern __shared__ unsigned int thread_status[2];
267
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> fc_d[2 * 1024];
268
+ T* factor_cache = reinterpret_cast<T*>(fc_d);
269
+
270
+ int orig_idx = blockDim.x * (blockIdx.x % n_blocks) + threadIdx.x;
271
+
272
+ int num_row = blockIdx.x / n_blocks;
273
+ int idx = 2 * orig_idx + 1;
274
+ const int k = 2;
275
+
276
+ if(idx >= n) {
277
+ return;
278
+ }
279
+
280
+ int group_num = idx / m;
281
+ int group_pos = idx % m;
282
+ T* out_off = out + num_row * n;
283
+ T* carries_off = carries + num_row * n_blocks * k;
284
+
285
+ T* group_start = out_off + m * group_num;
286
+ T* group_carries = carries_off + group_num * k;
287
+
288
+ const T* section_factors = factors;
289
+ T* section_carries = group_carries;
290
+
291
+ factor_cache[group_pos] = section_factors[group_pos];
292
+ factor_cache[group_pos - 1] = section_factors[group_pos - 1];
293
+ factor_cache[m + group_pos] = section_factors[m + group_pos];
294
+ factor_cache[m + group_pos - 1] = section_factors[m + group_pos - 1];
295
+ __syncthreads();
296
+
297
+ int pos = group_pos;
298
+ int up_bound = pos;
299
+ int low_bound = pos;
300
+ int rel_pos;
301
+
302
+ for(int level = 1, iter = 1; level < m; level *= 2, iter++) {
303
+ int sz = min(pow(2.0f, ((float) iter)), ((float) m));
304
+
305
+ if(level > 1) {
306
+ int factor = ceil(pos / ((float) sz));
307
+ up_bound = sz * factor - 1;
308
+ low_bound = up_bound - level + 1;
309
+ }
310
+
311
+ if(level == 1) {
312
+ pos = low_bound;
313
+ }
314
+
315
+ if(pos < low_bound) {
316
+ pos += level / 2;
317
+ }
318
+
319
+ if(pos + m * group_num >= n) {
320
+ break;
321
+ }
322
+
323
+ rel_pos = pos % level;
324
+ T carry = 0.0;
325
+ for(int i = 1; i <= min(k, level); i++) {
326
+ T k_value = group_start[low_bound - i];
327
+ const T* k_factors = factor_cache + m * (i - 1);
328
+ T factor = k_factors[rel_pos];
329
+ carry += k_value * factor;
330
+ }
331
+
332
+ group_start[pos] += carry;
333
+ __syncthreads();
334
+ }
335
+
336
+ if(pos >= m - k) {
337
+ if(carries != NULL) {
338
+ section_carries[pos - (m - k)] = group_start[pos];
339
+ }
340
+ }
341
+ }
342
+
343
+ template<typename T>
344
+ __global__ void correct_carries_sos(
345
+ const int m, const int n_blocks, const int carries_stride,
346
+ const int offset, const T* factors, T* carries) {
347
+
348
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> fcd3[4];
349
+ T* factor_cache = reinterpret_cast<T*>(fcd3);
350
+
351
+ int idx = threadIdx.x;
352
+ const int k = 2;
353
+ int pos = idx + (m - k);
354
+ T* row_carries = carries + carries_stride * blockIdx.x;
355
+
356
+ factor_cache[2 * idx] = factors[pos];
357
+ factor_cache[2 * idx + 1] = factors[m + pos];
358
+ __syncthreads();
359
+
360
+ for(int i = offset; i < n_blocks; i++) {
361
+ T* this_carries = row_carries + k * (i + (1 - offset));
362
+ T* prev_carries = row_carries + k * (i - offset);
363
+
364
+ T carry = 0.0;
365
+ for(int j = 1; j <= k; j++) {
366
+ // const T* k_factors = factors + m * (j - 1);
367
+ // T factor = k_factors[pos];
368
+ T factor = factor_cache[2 * idx + (j - 1)];
369
+ T k_value = prev_carries[k - j];
370
+ carry += factor * k_value;
371
+ }
372
+
373
+ this_carries[idx] += carry;
374
+ __syncthreads();
375
+ }
376
+ }
377
+
378
+ template<typename T>
379
+ __global__ void second_pass_iir_sos(
380
+ const int m, const int n, const int carries_stride,
381
+ const int n_blocks, const int offset, const T* factors,
382
+ T* carries, T* out) {
383
+
384
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> fcd2[2 * 1024];
385
+ T* factor_cache = reinterpret_cast<T*>(fcd2);
386
+
387
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> c_d[2];
388
+ T* carries_cache = reinterpret_cast<T*>(c_d);
389
+
390
+ int idx = blockDim.x * (blockIdx.x % n_blocks) + threadIdx.x;
391
+ idx += offset * m;
392
+
393
+ int row_num = blockIdx.x / n_blocks;
394
+ int n_group = idx / m;
395
+ int pos = idx % m;
396
+ const int k = 2;
397
+
398
+ T* out_off = out + row_num * n;
399
+ T* carries_off = carries + row_num * carries_stride;
400
+ const T* prev_carries = carries_off + (n_group - offset) * k;
401
+
402
+ if(pos < k) {
403
+ carries_cache[pos] = prev_carries[pos];
404
+ }
405
+
406
+ if(idx >= n) {
407
+ return;
408
+ }
409
+
410
+ factor_cache[pos] = factors[pos];
411
+ factor_cache[pos + m] = factors[pos + m];
412
+ __syncthreads();
413
+
414
+ T carry = 0.0;
415
+ for(int i = 1; i <= k; i++) {
416
+ const T* k_factors = factor_cache + m * (i - 1);
417
+ T factor = k_factors[pos];
418
+ T k_value = carries_cache[k - i];
419
+ carry += factor * k_value;
420
+ }
421
+
422
+ out_off[idx] += carry;
423
+ }
424
+
425
+ template<typename T>
426
+ __global__ void fir_sos(
427
+ const int m, const int n, const int carries_stride, const int n_blocks,
428
+ const int offset, const T* sos, T* carries, T* out) {
429
+
430
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> fir_cc[1024 + 2];
431
+ T* fir_cache = reinterpret_cast<T*>(fir_cc);
432
+
433
+ extern __shared__ __align__(sizeof(T)) thrust::complex<double> fir_b[3];
434
+ T* b = reinterpret_cast<T*>(fir_b);
435
+
436
+ int idx = blockDim.x * (blockIdx.x % n_blocks) + threadIdx.x;
437
+ int row_num = blockIdx.x / n_blocks;
438
+ int n_group = idx / m;
439
+ int pos = idx % m;
440
+ const int k = 2;
441
+
442
+ T* out_row = out + row_num * n;
443
+ T* out_off = out_row + n_group * m;
444
+ T* carries_off = carries + row_num * carries_stride;
445
+ T* this_carries = carries_off + k * (n_group + (1 - offset));
446
+ T* group_carries = carries_off + (n_group - offset) * k;
447
+
448
+ if(pos <= k) {
449
+ b[pos] = sos[pos];
450
+ }
451
+
452
+ if(pos < k) {
453
+ if(offset && n_group == 0) {
454
+ fir_cache[pos] = 0;
455
+ } else {
456
+ fir_cache[pos] = group_carries[pos];
457
+ }
458
+ }
459
+
460
+ if(idx >= n) {
461
+ return;
462
+ }
463
+
464
+ fir_cache[pos + k] = out_off[pos];
465
+ __syncthreads();
466
+
467
+ T acc = 0.0;
468
+ for(int i = k; i >= 0; i--) {
469
+ acc += fir_cache[pos + i] * b[k - i];
470
+ }
471
+
472
+ out_off[pos] = acc;
473
+ }
474
+ """ # NOQA
475
+
476
+ IIR_MODULE = cupy.RawModule(
477
+ code=IIR_KERNEL, options=('-std=c++11',),
478
+ name_expressions=[f'compute_correction_factors<{x}, {y}>'
479
+ for x, y in TYPE_PAIR_NAMES] +
480
+ [f'correct_carries<{x}>' for x in TYPE_NAMES] +
481
+ [f'first_pass_iir<{x}>' for x in TYPE_NAMES] +
482
+ [f'second_pass_iir<{x}>' for x in TYPE_NAMES])
483
+
484
+ IIR_SOS_MODULE = cupy.RawModule(
485
+ code=IIR_SOS_KERNEL, options=('-std=c++11',),
486
+ name_expressions=[f'compute_correction_factors_sos<{x}, {y}>'
487
+ for x, y in TYPE_PAIR_NAMES] +
488
+ [f'pick_carries<{x}>' for x in TYPE_NAMES] +
489
+ [f'correct_carries_sos<{x}>' for x in TYPE_NAMES] +
490
+ [f'first_pass_iir_sos<{x}>' for x in TYPE_NAMES] +
491
+ [f'second_pass_iir_sos<{x}>' for x in TYPE_NAMES] +
492
+ [f'fir_sos<{x}>' for x in TYPE_NAMES])
493
+
494
+
495
+ def _get_module_func(module, func_name, *template_args):
496
+ args_dtypes = [_get_typename(arg.dtype) for arg in template_args]
497
+ template = ', '.join(args_dtypes)
498
+ kernel_name = f'{func_name}<{template}>' if template_args else func_name
499
+ kernel = module.get_function(kernel_name)
500
+ return kernel
501
+
502
+
503
+ def collapse_2d(x, axis):
504
+ x = cupy.moveaxis(x, axis, -1)
505
+ x_shape = x.shape
506
+ x = x.reshape(-1, x.shape[-1])
507
+ if not x.flags.c_contiguous:
508
+ x = x.copy()
509
+ return x, x_shape
510
+
511
+
512
+ def collapse_2d_rest(x, axis):
513
+ x = cupy.moveaxis(x, axis + 1, -1)
514
+ x_shape = x.shape
515
+ x = x.reshape(x.shape[0], -1, x.shape[-1])
516
+ if not x.flags.c_contiguous:
517
+ x = x.copy()
518
+ return x, x_shape
519
+
520
+
521
+ def compute_correction_factors(a, block_sz, dtype):
522
+ k = a.size
523
+ correction = cupy.eye(k, dtype=dtype)
524
+ correction = cupy.c_[
525
+ correction[::-1], cupy.empty((k, block_sz), dtype=dtype)]
526
+ corr_kernel = _get_module_func(
527
+ IIR_MODULE, 'compute_correction_factors', correction, a)
528
+ corr_kernel((k,), (1,), (block_sz, k, a, correction))
529
+ return correction
530
+
531
+
532
+ def apply_iir(x, a, axis=-1, zi=None, dtype=None, block_sz=1024):
533
+ # GPU throughput is faster when using single precision floating point
534
+ # numbers
535
+ # x = x.astype(cupy.float32)
536
+ if dtype is None:
537
+ dtype = cupy.result_type(x.dtype, a.dtype)
538
+
539
+ a = a.astype(dtype)
540
+
541
+ if zi is not None:
542
+ zi = zi.astype(dtype)
543
+
544
+ x_shape = x.shape
545
+ x_ndim = x.ndim
546
+ axis = _normalize_axis_index(axis, x_ndim)
547
+ k = a.size
548
+ n = x_shape[axis]
549
+
550
+ if x_ndim > 1:
551
+ x, x_shape = collapse_2d(x, axis)
552
+ if zi is not None:
553
+ zi, _ = collapse_2d(zi, axis)
554
+
555
+ out = cupy.array(x, dtype=dtype, copy=True)
556
+
557
+ num_rows = 1 if x.ndim == 1 else x.shape[0]
558
+ n_blocks = (n + block_sz - 1) // block_sz
559
+ total_blocks = num_rows * n_blocks
560
+
561
+ correction = cupy.eye(k, dtype=dtype)
562
+ correction = cupy.c_[
563
+ correction[::-1], cupy.empty((k, block_sz), dtype=dtype)]
564
+ carries = cupy.empty(
565
+ (num_rows, n_blocks, k), dtype=dtype)
566
+
567
+ corr_kernel = _get_module_func(
568
+ IIR_MODULE, 'compute_correction_factors', correction, a)
569
+ first_pass_kernel = _get_module_func(IIR_MODULE, 'first_pass_iir', out)
570
+ second_pass_kernel = _get_module_func(IIR_MODULE, 'second_pass_iir', out)
571
+ carry_correction_kernel = _get_module_func(
572
+ IIR_MODULE, 'correct_carries', out)
573
+
574
+ corr_kernel((k,), (1,), (block_sz, k, a, correction))
575
+ first_pass_kernel((total_blocks,), (block_sz // 2,),
576
+ (block_sz, k, n, n_blocks, (n_blocks) * k,
577
+ correction, out, carries))
578
+
579
+ if zi is not None:
580
+ if zi.ndim == 1:
581
+ zi = cupy.broadcast_to(zi, (num_rows, 1, zi.size))
582
+ elif zi.ndim == 2:
583
+ zi = zi.reshape(num_rows, 1, zi.shape[-1])
584
+
585
+ if carries.size == 0:
586
+ carries = zi
587
+ else:
588
+ carries = cupy.concatenate((zi, carries), axis=1)
589
+
590
+ if not carries.flags.c_contiguous:
591
+ carries = carries.copy()
592
+
593
+ if n_blocks > 1 or zi is not None:
594
+ starting_group = int(zi is None)
595
+ blocks_to_merge = n_blocks - starting_group
596
+ carries_stride = (n_blocks + (1 - starting_group)) * k
597
+ carry_correction_kernel(
598
+ (num_rows,), (k,),
599
+ (block_sz, k, n_blocks, carries_stride, starting_group,
600
+ correction, carries))
601
+ second_pass_kernel(
602
+ (num_rows * blocks_to_merge,), (block_sz,),
603
+ (block_sz, k, n, carries_stride, blocks_to_merge,
604
+ starting_group, correction, carries, out))
605
+
606
+ if x_ndim > 1:
607
+ out = out.reshape(x_shape)
608
+ out = cupy.moveaxis(out, -1, axis)
609
+ if not out.flags.c_contiguous:
610
+ out = out.copy()
611
+
612
+ return out
613
+
614
+
615
+ def compute_correction_factors_sos(sos, block_sz, dtype):
616
+ n_sections = sos.shape[0]
617
+ correction = cupy.empty((n_sections, 2, block_sz), dtype=dtype)
618
+ corr_kernel = _get_module_func(
619
+ IIR_SOS_MODULE, 'compute_correction_factors_sos', correction, sos)
620
+ corr_kernel((n_sections,), (2,), (block_sz, sos, correction))
621
+ return correction
622
+
623
+
624
+ def apply_iir_sos(x, sos, axis=-1, zi=None, dtype=None, block_sz=1024,
625
+ apply_fir=True, out=None):
626
+ if dtype is None:
627
+ dtype = cupy.result_type(x.dtype, sos.dtype)
628
+
629
+ sos = sos.astype(dtype)
630
+
631
+ if zi is not None:
632
+ zi = zi.astype(dtype)
633
+
634
+ x_shape = x.shape
635
+ x_ndim = x.ndim
636
+ n_sections = sos.shape[0]
637
+ axis = _normalize_axis_index(axis, x_ndim)
638
+ k = 2
639
+ n = x_shape[axis]
640
+ zi_shape = None
641
+
642
+ if x_ndim > 1:
643
+ x, x_shape = collapse_2d(x, axis)
644
+
645
+ if zi is not None:
646
+ zi, zi_shape = collapse_2d_rest(zi, axis)
647
+
648
+ if out is None:
649
+ out = cupy.array(x, dtype=dtype, copy=True)
650
+
651
+ num_rows = 1 if x.ndim == 1 else x.shape[0]
652
+ n_blocks = (n + block_sz - 1) // block_sz
653
+ total_blocks = num_rows * n_blocks
654
+
655
+ correction = compute_correction_factors_sos(sos, block_sz, dtype)
656
+ carries = cupy.empty(
657
+ (num_rows, n_blocks, k), dtype=dtype)
658
+ all_carries = carries
659
+ zi_out = None
660
+ if zi is not None:
661
+ zi_out = cupy.empty_like(zi)
662
+ all_carries = cupy.empty(
663
+ (num_rows, n_blocks + 1, k), dtype=dtype)
664
+
665
+ first_pass_kernel = _get_module_func(
666
+ IIR_SOS_MODULE, 'first_pass_iir_sos', out)
667
+ second_pass_kernel = _get_module_func(
668
+ IIR_SOS_MODULE, 'second_pass_iir_sos', out)
669
+ carry_correction_kernel = _get_module_func(
670
+ IIR_SOS_MODULE, 'correct_carries_sos', out)
671
+ fir_kernel = _get_module_func(IIR_SOS_MODULE, 'fir_sos', out)
672
+ carries_kernel = _get_module_func(IIR_SOS_MODULE, 'pick_carries', out)
673
+
674
+ starting_group = int(zi is None)
675
+ blocks_to_merge = n_blocks - starting_group
676
+ carries_stride = (n_blocks + (1 - starting_group)) * k
677
+
678
+ carries_kernel((num_rows * n_blocks,), (k,),
679
+ (block_sz, n, carries_stride, n_blocks, starting_group,
680
+ out, all_carries))
681
+
682
+ for s in range(n_sections):
683
+ b = sos[s]
684
+ if zi is not None:
685
+ section_zi = zi[s, :, :2]
686
+ all_carries[:, 0, :] = section_zi
687
+ zi_out[s, :, :2] = axis_slice(out, n - 2, n)
688
+
689
+ if apply_fir:
690
+ fir_kernel((num_rows * n_blocks,), (block_sz,),
691
+ (block_sz, n, carries_stride, n_blocks, starting_group,
692
+ b, all_carries, out))
693
+
694
+ first_pass_kernel(
695
+ (total_blocks,), (block_sz // 2,),
696
+ (block_sz, n, n_blocks, correction[s], out, carries))
697
+
698
+ if n_blocks > 1 or zi is not None:
699
+ if zi is not None:
700
+ section_zi = zi[s, :, 2:]
701
+ all_carries[:, 0, :] = section_zi
702
+ all_carries[:, 1:, :] = carries
703
+
704
+ carry_correction_kernel(
705
+ (num_rows,), (k,),
706
+ (block_sz, n_blocks, carries_stride, starting_group,
707
+ correction[s], all_carries))
708
+ second_pass_kernel(
709
+ (num_rows * blocks_to_merge,), (block_sz,),
710
+ (block_sz, n, carries_stride, blocks_to_merge,
711
+ starting_group, correction[s], all_carries, out))
712
+
713
+ if apply_fir:
714
+ carries_kernel(
715
+ (num_rows * n_blocks,), (k,),
716
+ (block_sz, n, carries_stride, n_blocks, starting_group,
717
+ out, all_carries))
718
+
719
+ if zi is not None:
720
+ zi_out[s, :, 2:] = axis_slice(out, n - 2, n)
721
+
722
+ if x_ndim > 1:
723
+ out = out.reshape(x_shape)
724
+ out = cupy.moveaxis(out, -1, axis)
725
+ if not out.flags.c_contiguous:
726
+ out = out.copy()
727
+
728
+ if zi is not None:
729
+ zi_out = zi_out.reshape(zi_shape)
730
+ if len(zi_shape) > 2:
731
+ zi_out = cupy.moveaxis(zi_out, -1, axis)
732
+ if not zi_out.flags.c_contiguous:
733
+ zi_out = zi_out.copy()
734
+
735
+ if zi is not None:
736
+ return out, zi_out
737
+ return out
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_lti_conversion.py ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import cupy
2
+
3
+
4
+ def _none_to_empty_2d(arg):
5
+ if arg is None:
6
+ return cupy.zeros((0, 0))
7
+ else:
8
+ return arg
9
+
10
+
11
+ def _atleast_2d_or_none(arg):
12
+ if arg is not None:
13
+ return cupy.atleast_2d(arg)
14
+
15
+
16
+ def _shape_or_none(M):
17
+ if M is not None:
18
+ return M.shape
19
+ else:
20
+ return (None,) * 2
21
+
22
+
23
+ def _choice_not_none(*args):
24
+ for arg in args:
25
+ if arg is not None:
26
+ return arg
27
+
28
+
29
+ def _restore(M, shape):
30
+ if M.shape == (0, 0):
31
+ return cupy.zeros(shape)
32
+ else:
33
+ if M.shape != shape:
34
+ raise ValueError("The input arrays have incompatible shapes.")
35
+ return M
36
+
37
+
38
+ def abcd_normalize(A=None, B=None, C=None, D=None):
39
+ """Check state-space matrices and ensure they are 2-D.
40
+
41
+ If enough information on the system is provided, that is, enough
42
+ properly-shaped arrays are passed to the function, the missing ones
43
+ are built from this information, ensuring the correct number of
44
+ rows and columns. Otherwise a ValueError is raised.
45
+
46
+ Parameters
47
+ ----------
48
+ A, B, C, D : array_like, optional
49
+ State-space matrices. All of them are None (missing) by default.
50
+ See `ss2tf` for format.
51
+
52
+ Returns
53
+ -------
54
+ A, B, C, D : array
55
+ Properly shaped state-space matrices.
56
+
57
+ Raises
58
+ ------
59
+ ValueError
60
+ If not enough information on the system was provided.
61
+
62
+ """
63
+ A, B, C, D = map(_atleast_2d_or_none, (A, B, C, D))
64
+
65
+ MA, NA = _shape_or_none(A)
66
+ MB, NB = _shape_or_none(B)
67
+ MC, NC = _shape_or_none(C)
68
+ MD, ND = _shape_or_none(D)
69
+
70
+ p = _choice_not_none(MA, MB, NC)
71
+ q = _choice_not_none(NB, ND)
72
+ r = _choice_not_none(MC, MD)
73
+ if p is None or q is None or r is None:
74
+ raise ValueError("Not enough information on the system.")
75
+
76
+ A, B, C, D = map(_none_to_empty_2d, (A, B, C, D))
77
+ A = _restore(A, (p, p))
78
+ B = _restore(B, (p, q))
79
+ C = _restore(C, (r, p))
80
+ D = _restore(D, (r, q))
81
+
82
+ return A, B, C, D
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_ltisys.py ADDED
The diff for this file is too large to render. See raw diff
 
vllm/lib/python3.10/site-packages/cupyx/scipy/signal/_max_len_seq.py ADDED
@@ -0,0 +1,119 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import cupy
2
+
3
+
4
+ _mls_taps = {2: [1], 3: [2], 4: [3], 5: [3], 6: [5], 7: [6], 8: [7, 6, 1],
5
+ 9: [5], 10: [7], 11: [9], 12: [11, 10, 4], 13: [12, 11, 8],
6
+ 14: [13, 12, 2], 15: [14], 16: [15, 13, 4], 17: [14],
7
+ 18: [11], 19: [18, 17, 14], 20: [17], 21: [19], 22: [21],
8
+ 23: [18], 24: [23, 22, 17], 25: [22], 26: [25, 24, 20],
9
+ 27: [26, 25, 22], 28: [25], 29: [27], 30: [29, 28, 7],
10
+ 31: [28], 32: [31, 30, 10]}
11
+
12
+
13
+ MAX_LEN_SEQ_KERNEL = r"""
14
+ #include <cupy/math_constants.h>
15
+ #include <cupy/carray.cuh>
16
+ #include <cupy/complex.cuh>
17
+
18
+ extern "C" __global__ void max_len_seq(
19
+ long long length, long long n_taps, int n_state, long long* taps,
20
+ signed char* state, signed char* seq) {
21
+
22
+ int idx = blockDim.x * blockIdx.x + threadIdx.x;
23
+ for(long long i = 0; i < length; i++) {
24
+ signed char next_state = state[(idx + 1) % n_state];
25
+ if(idx == n_state - 1) {
26
+ seq[i] = state[0];
27
+ for(int n_tap = 0; n_tap < n_taps; n_tap++) {
28
+ long long tap = taps[n_tap];
29
+ next_state ^= state[tap];
30
+ }
31
+ }
32
+ state[idx] = next_state;
33
+ }
34
+ }
35
+ """
36
+
37
+ _max_len_seq = cupy.RawKernel(MAX_LEN_SEQ_KERNEL, 'max_len_seq')
38
+
39
+
40
+ def max_len_seq(nbits, state=None, length=None, taps=None):
41
+ """
42
+ Maximum length sequence (MLS) generator.
43
+
44
+ Parameters
45
+ ----------
46
+ nbits : int
47
+ Number of bits to use. Length of the resulting sequence will
48
+ be ``(2**nbits) - 1``. Note that generating long sequences
49
+ (e.g., greater than ``nbits == 16``) can take a long time.
50
+ state : array_like, optional
51
+ If array, must be of length ``nbits``, and will be cast to binary
52
+ (bool) representation. If None, a seed of ones will be used,
53
+ producing a repeatable representation. If ``state`` is all
54
+ zeros, an error is raised as this is invalid. Default: None.
55
+ length : int, optional
56
+ Number of samples to compute. If None, the entire length
57
+ ``(2**nbits) - 1`` is computed.
58
+ taps : array_like, optional
59
+ Polynomial taps to use (e.g., ``[7, 6, 1]`` for an 8-bit sequence).
60
+ If None, taps will be automatically selected (for up to
61
+ ``nbits == 32``).
62
+
63
+ Returns
64
+ -------
65
+ seq : array
66
+ Resulting MLS sequence of 0's and 1's.
67
+ state : array
68
+ The final state of the shift register.
69
+
70
+ Notes
71
+ -----
72
+ The algorithm for MLS generation is generically described in:
73
+
74
+ https://en.wikipedia.org/wiki/Maximum_length_sequence
75
+
76
+ The default values for taps are specifically taken from the first
77
+ option listed for each value of ``nbits`` in:
78
+
79
+ https://web.archive.org/web/20181001062252/http://www.newwaveinstruments.com/resources/articles/m_sequence_linear_feedback_shift_register_lfsr.htm
80
+
81
+ """ # NOQA
82
+ if taps is None:
83
+ if nbits not in _mls_taps:
84
+ known_taps = cupy.array(list(_mls_taps.keys()))
85
+ raise ValueError('nbits must be between %s and %s if taps is None'
86
+ % (known_taps.min(), known_taps.max()))
87
+ taps = cupy.array(_mls_taps[nbits], cupy.int64)
88
+ else:
89
+ taps = cupy.unique(cupy.array(taps, cupy.int64))[::-1]
90
+ if cupy.any(taps < 0) or cupy.any(taps > nbits) or taps.size < 1:
91
+ raise ValueError('taps must be non-empty with values between '
92
+ 'zero and nbits (inclusive)')
93
+ taps = cupy.array(taps) # needed for Cython and Pythran
94
+
95
+ n_max = (2 ** nbits) - 1
96
+ if length is None:
97
+ length = n_max
98
+ else:
99
+ length = int(length)
100
+ if length < 0:
101
+ raise ValueError('length must be greater than or equal to 0')
102
+
103
+ # We use int8 instead of bool here because NumPy arrays of bools
104
+ # don't seem to work nicely with Cython
105
+ if state is None:
106
+ state = cupy.ones(nbits, dtype=cupy.int8, order='c')
107
+ else:
108
+ # makes a copy if need be, ensuring it's 0's and 1's
109
+ state = cupy.array(state, dtype=bool, order='c').astype(cupy.int8)
110
+ if state.ndim != 1 or state.size != nbits:
111
+ raise ValueError('state must be a 1-D array of size nbits')
112
+ if cupy.all(state == 0):
113
+ raise ValueError('state must not be all zeros')
114
+
115
+ seq = cupy.empty(length, dtype=cupy.int8, order='c')
116
+ n_taps = len(taps)
117
+
118
+ _max_len_seq((1,), (nbits,), (length, n_taps, nbits, taps, state, seq))
119
+ return seq, state