PyScalapack is a Python wrapper for ScaLAPACK.
To use PySCALAPACK, users must provide the path to the ScaLAPACK dynamic shared library, which is loaded by `ctypes.CDLL` by default.
# Install
Please either copy or create a soft link for the directory in the `site-packages` directory.
Alternatively, users can utilize pip to install the PyScalapack package by running the command `pip install PyScalapack`.
# Documents
## Load ScaLAPACK
The ScaLAPACK dynamic shared library should be loaded prior to engaging in any further operations.
import PyScalapack
scalapack = PyScalapack("libscalapack.so")
If the ScaLAPACK dynamic shared library reside outside the default path, users must supply their absolute paths.
In case ScaLAPACK functions are distributed across multiple cdecl convention dynamic shared libraries,
include them all when invoking `PySCALAPACK`. For instance, use `PySCALAPACK("libmkl_core.so", ...)`.
To override the default loader `ctypes.CDLL`, add a keyword-only argument called `loader` to `PySCALAPACK`,
which is particularly helpful when working with non-cdecl convention shared libraries.
## Create a context
Create a BLACS context to facilitate subsequent BLACS or ScaLAPACK operations.
A context informs ScaLAPACK on how a distributed matrix is positioned among various processes.
Establishing a context is mandatory for creating a BLACS matrix.
Only three parameters are required: layout (either column major or row major), the number of rows in the process grid, and the number of columns in the process grid.
Set the layout to b'C' for column major or b'R' for row major.
To ensure efficient use of resources and prevent idle processes, make sure the product of `nprow` and `npcol` is equal to the number of processes.
If the product of `nprow` and `npcol` surpasses the number of processes, a fatal error arises.
When the product of `nprow` and `npcol` is smaller than the number pf processes, some processes may be excluded from the context.
These excluded processes are marked as invalid within the context.
To check if the current process is valid, users can examine the `context.valid` attribute.
Alternatively, they can also utilize boolean operations such as `bool(context)`.
The context in PyScalapack has several attributes including:
- `layout`: layout of the process grid, it is either 'R' for row major or 'C' for column major;
- `nprow`: row number of the process grid;
- `npcol`: column number of the process grid;
- `rank`: rank of the current process;
- `size`: size of the process grid;
- `ictx`: the raw context handle from BLACS;
- `myrow`: row index of the current process;
- `mycol`: column index of the current process;
- `valid`: whether the current process is valid, it equals `rank < nprol * npcol`.
Most of these attributes are of type ctypes bool or ctypes int.
To obtain their Python values, users can access them using the `value` attribute, like `context.nprow.value`.
import PyScalapack
scalapack = PyScalapack("libscalapack.so")
with scalapack(layout=b'C', nprow=1, npcol=1) as context:
for key, value in context.__dict__.items():
print(f"{key} = {value}")
scalapack = <PyScalapack.Scalapack object at 0x7f6242fa0210>
layout = c_char(b'C')
nprow = c_int(1)
npcol = c_int(1)
rank = c_int(0)
size = c_int(1)
ictxt = c_int(0)
myrow = c_int(0)
mycol = c_int(0)
valid = True
Users can utilize the function `context.barrier(scope=b'A')` to synchronize all processes within the process grid.
Additionally, calling with `scope=b'R'` will synchronize all processes in the same row of the process grid,
while invoking `context.barrier` with `scope=b'C'` will synchronize all processes in the same column of the process grid.
## Create an array
Utilize `context.array` to generate a block-cyclic distributed array.
The matrix's shape relies on the arguments `m` and `n`, whereas the block size for distribution among processes is set by `mb` and `nb`.
Once an array is created, each process will have its own local matrix dimensions, which can be accessed through `local_m` and `local_n`.
import numpy as np
import PyScalapack
scalapack = PyScalapack(
"libmpi.so",
"libmkl_core.so",
"libmkl_sequential.so",
"libmkl_intel_lp64.so",
"libmkl_blacs_intelmpi_lp64.so",
"libmkl_scalapack_lp64.so",
)
with scalapack(b'C', 2, 2) as context:
array = context.array(
m=23,
n=47,
mb=5,
nb=5,
dtype=np.float64,
)
if context.rank.value == 0:
print(f"Matrix dimension is ({array.m}, {array.n})")
print(f"Matrix local dimension at process " + #
f"({context.myrow.value}, {context.mycol.value})" + #
f" is ({array.local_m}, {array.local_n})")
Matrix dimension is (23, 47)
Matrix local dimension at process (0, 0) is (13, 25)
Matrix local dimension at process (1, 0) is (10, 25)
Matrix local dimension at process (0, 1) is (13, 22)
Matrix local dimension at process (1, 1) is (10, 22)
The user can create a new empty matrix with the desired scalar type by specifying `dtype`.
Alternatively, they can provide an existing distributed matrix by passing local matrix to `data` argument,
making sure that the local dimensions of the matrix remains accurate across all processes.
Regardless of how the array was generated,
users can access the local matrix data by using `array.data`, and retrieve the scalar type via `array.dtype`.
import numpy as np
import PyScalapack
scalapack = PyScalapack("libscalapack.so")
with scalapack(b'C', 1, 1) as context:
array = context.array(
m=128,
n=512,
mb=1,
nb=1,
data=np.zeros([128, 512], order='F'),
)
print(f"Matrix dimension is ({array.m}, {array.n})")
print(f"Matrix local dimension is " + #
f"({array.local_m}, {array.local_n})")
with scalapack(b'R', 1, 1) as context:
array = context.array(
m=128,
n=512,
mb=1,
nb=1,
data=np.zeros([128, 512], order='C'),
)
print(f"Matrix dimension is ({array.m}, {array.n})")
print(f"Matrix local dimension is " + #
f"({array.local_m}, {array.local_n})")
Matrix dimension is (128, 512)
Matrix local dimension is (128, 512)
Matrix dimension is (128, 512)
Matrix local dimension is (128, 512)
When passing a given local matrix, make sure the NumPy array order matches the context layout.
Use `'F'` for column major layout and `'C'` for row major layout.
## Redistribute matrix
Within ScaLAPACK, the `p?gemr2d` subroutine serves as a tool for redistributing matrix.
To redistribute a matrix from one context to another with `p?gemr2d` in ScaLAPACK,
users should furnish the matrix's dimensions, details about both matrices (which can be acquired via `scalapack_params()`),
and one raw BLACS context handle to the subroutine.
import numpy as np
import PyScalapack
scalapack = PyScalapack(
"libmpi.so",
"libmkl_core.so",
"libmkl_sequential.so",
"libmkl_intel_lp64.so",
"libmkl_blacs_intelmpi_lp64.so",
"libmkl_scalapack_lp64.so",
)
with (
scalapack(b'C', 1, 2) as context1,
scalapack(b'C', 2, 1) as context2,
):
m = 2
n = 2
array1 = context1.array(m, n, 1, 1, dtype=np.float64)
array1.data[...] = np.random.randn(*array1.data.shape)
print(f"rank={context1.rank.value} before " + #
f"redistribute {array1.data.reshape([-1])}")
array2 = context2.array(m, n, 1, 1, dtype=np.float64)
scalapack.pgemr2d["D"](
*(m, n),
*array1.scalapack_params(),
*array2.scalapack_params(),
context1.ictxt,
)
print(f"rank={context2.rank.value} after " + #
f"redistribute {array2.data.reshape([-1])}")
rank=0 before redistribute [0.90707631 1.18754568]
rank=0 after redistribute [0.90707631 0.75556488]
rank=1 before redistribute [ 0.75556488 -0.4480556 ]
rank=1 after redistribute [ 1.18754568 -0.4480556 ]
## Call ScaLAPACK function
Here's an example that demonstrates calling pdgemm and comparing its result to a similar calculation performed by NumPy.
We create two contexts, `context` serves as the primary one while `context0` acts as a supplemental context containing solely rank-0 processes tailored for data redistribution.
Initially, we produce a random matrix within `context0` and redistribute it to `context`.
Post-redistribution, we invoke `pdgemm` to execute matrix multiplication within `context`.
Following this operation, we redistribute the resulting product back to `context0` and contrast it with the computation derived using NumPy.
import numpy as np
import PyScalapack
scalapack = PyScalapack(
"libmpi.so",
"libmkl_core.so",
"libmkl_sequential.so",
"libmkl_intel_lp64.so",
"libmkl_blacs_intelmpi_lp64.so",
"libmkl_scalapack_lp64.so",
)
L1 = 128
L2 = 512
with (
scalapack(b'C', 2, 2) as context,
scalapack(b'C', 1, 1) as context0,
):
array0 = context0.array(L1, L2, 1, 1, dtype=np.float64)
if context0:
array0.data[...] = np.random.randn(*array0.data.shape)
array = context.array(L1, L2, 1, 1, dtype=np.float64)
scalapack.pgemr2d["D"](
*(L1, L2),
*array0.scalapack_params(),
*array.scalapack_params(),
context.ictxt,
)
result = context.array(L1, L1, 1, 1, dtype=np.float64)
scalapack.pdgemm(
b'N',
b'T',
*(L1, L1, L2),
scalapack.d_one,
*array.scalapack_params(),
*array.scalapack_params(),
scalapack.d_zero,
*result.scalapack_params(),
)
result0 = context0.array(L1, L1, 1, 1, dtype=np.float64)
scalapack.pgemr2d["D"](
*(L1, L1),
*result.scalapack_params(),
*result0.scalapack_params(),
context.ictxt,
)
if context0:
error = result0.data - array0.data @ array0.data.T
print(np.linalg.norm(error))
2.931808596345247e-12
## Call LAPACK function
This package also offers a convenient interface for easily invoking LAPACK/BLAS functions.
The subsequent code demonstrates an instance of calling `dgemm`.
Users must additionally create an trivial context and create single-process ScaLAPACK array prior to invoking LAPACK/BLAS functions.
import numpy as np
import PyScalapack
scalapack = PyScalapack("libscalapack.so")
L1 = 128
L2 = 512
with scalapack(b'C', 1, 1) as context:
array = context.array(L1, L2, 1, 1, dtype=np.float64)
array.data[...] = np.random.randn(*array.data.shape)
result = context.array(L1, L1, 1, 1, dtype=np.float64)
scalapack.dgemm(
b'N',
b'T',
*(L1, L1, L2),
scalapack.d_one,
*array.lapack_params(),
*array.lapack_params(),
scalapack.d_zero,
*result.lapack_params(),
)
diff = result.data - array.data @ array.data.T
print(np.linalg.norm(diff))
0.0
## Generic variables and functions
As ScaLAPACK functions require scalar arguments of raw C types such as `c_int` or `c_float`,
we have defined several constant variables, including `zero = ctypes.c_int(0)`, `one = ctypes.c_int(1)`, and `neg_one = ctypes.c_int(-1)`.
The floating one and zero are also named as `?_one` and `?_zero`, where `?` represents `c`, `d`, `c` or `z`.
`f_one` and `f_zero` allow you to obtain the floating-point constant variables, depending on chosen scalar type.
import PyScalapack
scalapack = PyScalapack("libscalapack.so")
print(scalapack.f_one["D"] == scalapack.d_one)
print(scalapack.f_zero["Z"] == scalapack.z_zero)
True
True
Some functions like `p?gemm` can be chosen with `pgemm[char]`, where char represents `S`, `D`, `C` or `Z`.
But not all functions have this mapping because it's mapped manually based on our current needs.
Users can either map additional ScaLAPACK functions on their own, report issues, or submit pull requests.
import PyScalapack
scalapack = PyScalapack("libscalapack.so")
print(scalapack.pgemm["D"] == scalapack.pdgemm)
True
Raw data
{
"_id": null,
"home_page": null,
"name": "PyScalapack",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.7",
"maintainer_email": null,
"keywords": "scalapack, lapack, blas, linear algebra, scientific computing, parallel computing, MPI",
"author": null,
"author_email": "Hao Zhang <zh970205@mail.ustc.edu.cn>",
"download_url": null,
"platform": null,
"description": "PyScalapack is a Python wrapper for ScaLAPACK.\nTo use PySCALAPACK, users must provide the path to the ScaLAPACK dynamic shared library, which is loaded by `ctypes.CDLL` by default.\n\n\n# Install\n\nPlease either copy or create a soft link for the directory in the `site-packages` directory.\nAlternatively, users can utilize pip to install the PyScalapack package by running the command `pip install PyScalapack`.\n\n\n# Documents\n\n\n## Load ScaLAPACK\n\nThe ScaLAPACK dynamic shared library should be loaded prior to engaging in any further operations.\n\n import PyScalapack\n \n scalapack = PyScalapack(\"libscalapack.so\")\n\nIf the ScaLAPACK dynamic shared library reside outside the default path, users must supply their absolute paths.\nIn case ScaLAPACK functions are distributed across multiple cdecl convention dynamic shared libraries,\ninclude them all when invoking `PySCALAPACK`. For instance, use `PySCALAPACK(\"libmkl_core.so\", ...)`.\nTo override the default loader `ctypes.CDLL`, add a keyword-only argument called `loader` to `PySCALAPACK`,\nwhich is particularly helpful when working with non-cdecl convention shared libraries.\n\n\n## Create a context\n\nCreate a BLACS context to facilitate subsequent BLACS or ScaLAPACK operations.\nA context informs ScaLAPACK on how a distributed matrix is positioned among various processes.\nEstablishing a context is mandatory for creating a BLACS matrix.\nOnly three parameters are required: layout (either column major or row major), the number of rows in the process grid, and the number of columns in the process grid.\nSet the layout to b'C' for column major or b'R' for row major.\n\nTo ensure efficient use of resources and prevent idle processes, make sure the product of `nprow` and `npcol` is equal to the number of processes.\nIf the product of `nprow` and `npcol` surpasses the number of processes, a fatal error arises.\nWhen the product of `nprow` and `npcol` is smaller than the number pf processes, some processes may be excluded from the context.\nThese excluded processes are marked as invalid within the context.\nTo check if the current process is valid, users can examine the `context.valid` attribute.\nAlternatively, they can also utilize boolean operations such as `bool(context)`.\n\nThe context in PyScalapack has several attributes including:\n\n- `layout`: layout of the process grid, it is either 'R' for row major or 'C' for column major;\n- `nprow`: row number of the process grid;\n- `npcol`: column number of the process grid;\n- `rank`: rank of the current process;\n- `size`: size of the process grid;\n- `ictx`: the raw context handle from BLACS;\n- `myrow`: row index of the current process;\n- `mycol`: column index of the current process;\n- `valid`: whether the current process is valid, it equals `rank < nprol * npcol`.\n\nMost of these attributes are of type ctypes bool or ctypes int.\nTo obtain their Python values, users can access them using the `value` attribute, like `context.nprow.value`.\n\n import PyScalapack\n \n scalapack = PyScalapack(\"libscalapack.so\")\n \n with scalapack(layout=b'C', nprow=1, npcol=1) as context:\n for key, value in context.__dict__.items():\n \tprint(f\"{key} = {value}\")\n\n scalapack = <PyScalapack.Scalapack object at 0x7f6242fa0210>\n layout = c_char(b'C')\n nprow = c_int(1)\n npcol = c_int(1)\n rank = c_int(0)\n size = c_int(1)\n ictxt = c_int(0)\n myrow = c_int(0)\n mycol = c_int(0)\n valid = True\n\nUsers can utilize the function `context.barrier(scope=b'A')` to synchronize all processes within the process grid.\nAdditionally, calling with `scope=b'R'` will synchronize all processes in the same row of the process grid,\nwhile invoking `context.barrier` with `scope=b'C'` will synchronize all processes in the same column of the process grid.\n\n\n## Create an array\n\nUtilize `context.array` to generate a block-cyclic distributed array.\nThe matrix's shape relies on the arguments `m` and `n`, whereas the block size for distribution among processes is set by `mb` and `nb`.\nOnce an array is created, each process will have its own local matrix dimensions, which can be accessed through `local_m` and `local_n`.\n\n import numpy as np\n import PyScalapack\n \n scalapack = PyScalapack(\n \"libmpi.so\",\n \"libmkl_core.so\",\n \"libmkl_sequential.so\",\n \"libmkl_intel_lp64.so\",\n \"libmkl_blacs_intelmpi_lp64.so\",\n \"libmkl_scalapack_lp64.so\",\n )\n \n with scalapack(b'C', 2, 2) as context:\n array = context.array(\n \tm=23,\n \tn=47,\n \tmb=5,\n \tnb=5,\n \tdtype=np.float64,\n )\n if context.rank.value == 0:\n \tprint(f\"Matrix dimension is ({array.m}, {array.n})\")\n print(f\"Matrix local dimension at process \" + #\n \t f\"({context.myrow.value}, {context.mycol.value})\" + #\n \t f\" is ({array.local_m}, {array.local_n})\")\n\n Matrix dimension is (23, 47)\n Matrix local dimension at process (0, 0) is (13, 25)\n Matrix local dimension at process (1, 0) is (10, 25)\n Matrix local dimension at process (0, 1) is (13, 22)\n Matrix local dimension at process (1, 1) is (10, 22)\n\nThe user can create a new empty matrix with the desired scalar type by specifying `dtype`.\nAlternatively, they can provide an existing distributed matrix by passing local matrix to `data` argument,\nmaking sure that the local dimensions of the matrix remains accurate across all processes.\nRegardless of how the array was generated,\nusers can access the local matrix data by using `array.data`, and retrieve the scalar type via `array.dtype`.\n\n import numpy as np\n import PyScalapack\n \n scalapack = PyScalapack(\"libscalapack.so\")\n \n with scalapack(b'C', 1, 1) as context:\n array = context.array(\n \tm=128,\n \tn=512,\n \tmb=1,\n \tnb=1,\n \tdata=np.zeros([128, 512], order='F'),\n )\n print(f\"Matrix dimension is ({array.m}, {array.n})\")\n print(f\"Matrix local dimension is \" + #\n \t f\"({array.local_m}, {array.local_n})\")\n \n with scalapack(b'R', 1, 1) as context:\n array = context.array(\n \tm=128,\n \tn=512,\n \tmb=1,\n \tnb=1,\n \tdata=np.zeros([128, 512], order='C'),\n )\n print(f\"Matrix dimension is ({array.m}, {array.n})\")\n print(f\"Matrix local dimension is \" + #\n \t f\"({array.local_m}, {array.local_n})\")\n\n Matrix dimension is (128, 512)\n Matrix local dimension is (128, 512)\n Matrix dimension is (128, 512)\n Matrix local dimension is (128, 512)\n\nWhen passing a given local matrix, make sure the NumPy array order matches the context layout.\nUse `'F'` for column major layout and `'C'` for row major layout.\n\n\n## Redistribute matrix\n\nWithin ScaLAPACK, the `p?gemr2d` subroutine serves as a tool for redistributing matrix.\nTo redistribute a matrix from one context to another with `p?gemr2d` in ScaLAPACK,\nusers should furnish the matrix's dimensions, details about both matrices (which can be acquired via `scalapack_params()`),\nand one raw BLACS context handle to the subroutine.\n\n import numpy as np\n import PyScalapack\n \n scalapack = PyScalapack(\n \"libmpi.so\",\n \"libmkl_core.so\",\n \"libmkl_sequential.so\",\n \"libmkl_intel_lp64.so\",\n \"libmkl_blacs_intelmpi_lp64.so\",\n \"libmkl_scalapack_lp64.so\",\n )\n \n with (\n \tscalapack(b'C', 1, 2) as context1,\n \tscalapack(b'C', 2, 1) as context2,\n ):\n m = 2\n n = 2\n array1 = context1.array(m, n, 1, 1, dtype=np.float64)\n array1.data[...] = np.random.randn(*array1.data.shape)\n print(f\"rank={context1.rank.value} before \" + #\n \t f\"redistribute {array1.data.reshape([-1])}\")\n array2 = context2.array(m, n, 1, 1, dtype=np.float64)\n scalapack.pgemr2d[\"D\"](\n \t*(m, n),\n \t*array1.scalapack_params(),\n \t*array2.scalapack_params(),\n \tcontext1.ictxt,\n )\n print(f\"rank={context2.rank.value} after \" + #\n \t f\"redistribute {array2.data.reshape([-1])}\")\n\n rank=0 before redistribute [0.90707631 1.18754568]\n rank=0 after redistribute [0.90707631 0.75556488]\n rank=1 before redistribute [ 0.75556488 -0.4480556 ]\n rank=1 after redistribute [ 1.18754568 -0.4480556 ]\n\n\n## Call ScaLAPACK function\n\nHere's an example that demonstrates calling pdgemm and comparing its result to a similar calculation performed by NumPy.\nWe create two contexts, `context` serves as the primary one while `context0` acts as a supplemental context containing solely rank-0 processes tailored for data redistribution.\nInitially, we produce a random matrix within `context0` and redistribute it to `context`.\nPost-redistribution, we invoke `pdgemm` to execute matrix multiplication within `context`.\nFollowing this operation, we redistribute the resulting product back to `context0` and contrast it with the computation derived using NumPy.\n\n import numpy as np\n import PyScalapack\n \n scalapack = PyScalapack(\n \"libmpi.so\",\n \"libmkl_core.so\",\n \"libmkl_sequential.so\",\n \"libmkl_intel_lp64.so\",\n \"libmkl_blacs_intelmpi_lp64.so\",\n \"libmkl_scalapack_lp64.so\",\n )\n \n L1 = 128\n L2 = 512\n with (\n \tscalapack(b'C', 2, 2) as context,\n \tscalapack(b'C', 1, 1) as context0,\n ):\n array0 = context0.array(L1, L2, 1, 1, dtype=np.float64)\n if context0:\n \tarray0.data[...] = np.random.randn(*array0.data.shape)\n \n array = context.array(L1, L2, 1, 1, dtype=np.float64)\n scalapack.pgemr2d[\"D\"](\n \t*(L1, L2),\n \t*array0.scalapack_params(),\n \t*array.scalapack_params(),\n \tcontext.ictxt,\n )\n \n result = context.array(L1, L1, 1, 1, dtype=np.float64)\n scalapack.pdgemm(\n \tb'N',\n \tb'T',\n \t*(L1, L1, L2),\n \tscalapack.d_one,\n \t*array.scalapack_params(),\n \t*array.scalapack_params(),\n \tscalapack.d_zero,\n \t*result.scalapack_params(),\n )\n \n result0 = context0.array(L1, L1, 1, 1, dtype=np.float64)\n scalapack.pgemr2d[\"D\"](\n \t*(L1, L1),\n \t*result.scalapack_params(),\n \t*result0.scalapack_params(),\n \tcontext.ictxt,\n )\n \n if context0:\n \terror = result0.data - array0.data @ array0.data.T\n \tprint(np.linalg.norm(error))\n\n 2.931808596345247e-12\n\n\n## Call LAPACK function\n\nThis package also offers a convenient interface for easily invoking LAPACK/BLAS functions.\nThe subsequent code demonstrates an instance of calling `dgemm`.\nUsers must additionally create an trivial context and create single-process ScaLAPACK array prior to invoking LAPACK/BLAS functions.\n\n import numpy as np\n import PyScalapack\n \n scalapack = PyScalapack(\"libscalapack.so\")\n \n L1 = 128\n L2 = 512\n with scalapack(b'C', 1, 1) as context:\n array = context.array(L1, L2, 1, 1, dtype=np.float64)\n array.data[...] = np.random.randn(*array.data.shape)\n \n result = context.array(L1, L1, 1, 1, dtype=np.float64)\n scalapack.dgemm(\n \tb'N',\n \tb'T',\n \t*(L1, L1, L2),\n \tscalapack.d_one,\n \t*array.lapack_params(),\n \t*array.lapack_params(),\n \tscalapack.d_zero,\n \t*result.lapack_params(),\n )\n \n diff = result.data - array.data @ array.data.T\n print(np.linalg.norm(diff))\n\n 0.0\n\n\n## Generic variables and functions\n\nAs ScaLAPACK functions require scalar arguments of raw C types such as `c_int` or `c_float`,\nwe have defined several constant variables, including `zero = ctypes.c_int(0)`, `one = ctypes.c_int(1)`, and `neg_one = ctypes.c_int(-1)`.\nThe floating one and zero are also named as `?_one` and `?_zero`, where `?` represents `c`, `d`, `c` or `z`.\n`f_one` and `f_zero` allow you to obtain the floating-point constant variables, depending on chosen scalar type.\n\n import PyScalapack\n \n scalapack = PyScalapack(\"libscalapack.so\")\n \n print(scalapack.f_one[\"D\"] == scalapack.d_one)\n print(scalapack.f_zero[\"Z\"] == scalapack.z_zero)\n\n True\n True\n\nSome functions like `p?gemm` can be chosen with `pgemm[char]`, where char represents `S`, `D`, `C` or `Z`.\nBut not all functions have this mapping because it's mapped manually based on our current needs.\nUsers can either map additional ScaLAPACK functions on their own, report issues, or submit pull requests.\n\n import PyScalapack\n \n scalapack = PyScalapack(\"libscalapack.so\")\n \n print(scalapack.pgemm[\"D\"] == scalapack.pdgemm)\n\n True\n\n",
"bugtrack_url": null,
"license": "GPLv3",
"summary": "python wrapper for scalapack",
"version": "0.3.17",
"project_urls": {
"Changelog": "https://github.com/USTC-TNS/TNSP/blob/main/CHANGELOG.org",
"Homepage": "https://github.com/USTC-TNS/TNSP/tree/main/PyScalapack",
"Issues": "https://github.com/USTC-TNS/TNSP/issues",
"Repository": "https://github.com/USTC-TNS/TNSP.git"
},
"split_keywords": [
"scalapack",
" lapack",
" blas",
" linear algebra",
" scientific computing",
" parallel computing",
" mpi"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "384f52c72c24da2244032d4906b77587bd55a2ce43f6c8e93ac0987bf8fde9aa",
"md5": "b2ee583d16542e1d078cc1c292d79ec2",
"sha256": "688181bbda974545a089e1ae88a368def84cde2434b611991d459b0997aa17d3"
},
"downloads": -1,
"filename": "PyScalapack-0.3.17-py3-none-any.whl",
"has_sig": false,
"md5_digest": "b2ee583d16542e1d078cc1c292d79ec2",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.7",
"size": 9421,
"upload_time": "2024-09-06T07:35:08",
"upload_time_iso_8601": "2024-09-06T07:35:08.330891Z",
"url": "https://files.pythonhosted.org/packages/38/4f/52c72c24da2244032d4906b77587bd55a2ce43f6c8e93ac0987bf8fde9aa/PyScalapack-0.3.17-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-09-06 07:35:08",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "USTC-TNS",
"github_project": "TNSP",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "pyscalapack"
}
JSON
