Rename cb2 to evenodd.

Bug fix for GPT interface.

pyquda_core/pyquda/__init__.py

@@ -6,7 +6,6 @@
 from mpi4py import MPI
 
 from ._version import __version__  # noqa: F401
-from . import pyquda as quda
 from .field import LatticeInfo
 
 
@@ -141,7 +140,7 @@ def _getDefaultGrid(mpi_size: int, latt_size: List[int]):
     return min(min_grid)
 
 
-def _initEnviron(**kwargs):
+def _setEnviron(**kwargs):
     def _setEnviron(env, key, value):
         if value is not None:
             if env in environ:
@@ -154,7 +153,7 @@ def _setEnviron(env, key, value):
         _setEnviron(f"QUDA_{key.upper()}", key, kwargs[key])
 
 
-def _initEnvironWarn(**kwargs):
+def _setEnvironWarn(**kwargs):
     def _setEnviron(env, key, value):
         if value is not None:
             if env in environ:
@@ -172,9 +171,8 @@ def _setEnviron(env, key, value):
 
 def initGPU(backend: Literal["numpy", "cupy", "torch"] = None, gpuid: int = -1):
     global _CUDA_BACKEND, _HIP, _GPUID, _COMPUTE_CAPABILITY
-
     if isGridInitialized():
-        _MPI_LOGGER.critical("initGPU should be called before init", RuntimeError)
+        _MPI_LOGGER.critical("initGPU should be called before initGrid", RuntimeError)
     if _GPUID < 0:
         from platform import node as gethostname
 
@@ -239,17 +237,31 @@ def initGPU(backend: Literal["numpy", "cupy", "torch"] = None, gpuid: int = -1):
         _MPI_LOGGER.warning("GPU is already initialized", RuntimeWarning)
 
 
-def initQUDA(grid_size: List[int], gpuid: int):
+def initGrid(grid_size: List[int]):
+    global _GRID_SIZE, _GRID_COORD
+    if _GRID_SIZE is None:
+        Gx, Gy, Gz, Gt = grid_size
+        if _MPI_SIZE != Gx * Gy * Gz * Gt:
+            _MPI_LOGGER.critical(f"The MPI size {_MPI_SIZE} does not match the grid size {grid_size}", ValueError)
+        _GRID_SIZE = [Gx, Gy, Gz, Gt]
+        _GRID_COORD = getCoordFromRank(_MPI_RANK, _GRID_SIZE)
+        _MPI_LOGGER.info(f"Using the grid size {_GRID_SIZE}")
+    else:
+        _MPI_LOGGER.warning("Grid is already initialized", RuntimeWarning)
+
+
+def initQUDA(grid_size: List[int], gpuid: int, use_quda_allocator: bool = False):
     import atexit
+    from . import pyquda as quda, malloc_pyquda
 
-    # if _CUDA_BACKEND == "cupy":
-    #     import cupy
-    #     from . import malloc_pyquda
+    if use_quda_allocator:
+        if _CUDA_BACKEND == "cupy":
+            import cupy
 
-    #     allocator = cupy.cuda.PythonFunctionAllocator(
-    #         malloc_pyquda.pyquda_device_malloc, malloc_pyquda.pyquda_device_free
-    #     )
-    #     cupy.cuda.set_allocator(allocator.malloc)
+            allocator = cupy.cuda.PythonFunctionAllocator(
+                malloc_pyquda.pyquda_device_malloc, malloc_pyquda.pyquda_device_free
+            )
+            cupy.cuda.set_allocator(allocator.malloc)
 
     quda.initCommsGridQuda(4, grid_size)
     quda.initQuda(gpuid)
@@ -293,28 +305,23 @@ def init(
     """
     Initialize MPI along with the QUDA library.
     """
-    global _GRID_SIZE, _GRID_COORD, _DEFAULT_LATTICE
+    global _DEFAULT_LATTICE
     if _GRID_SIZE is None:
         initGPU(backend)
 
         use_default_grid = grid_size is None and latt_size is not None
         use_default_latt = latt_size is not None and t_boundary is not None and anisotropy is not None
         if use_default_grid:
             grid_size = _getDefaultGrid(_MPI_SIZE, latt_size)
-        Gx, Gy, Gz, Gt = grid_size if grid_size is not None else [1, 1, 1, 1]
-        if _MPI_SIZE != Gx * Gy * Gz * Gt:
-            _MPI_LOGGER.critical(f"The MPI size {_MPI_SIZE} does not match the grid size {grid_size}", ValueError)
-        _GRID_SIZE = [Gx, Gy, Gz, Gt]
-        _GRID_COORD = getCoordFromRank(_MPI_RANK, _GRID_SIZE)
-        _MPI_LOGGER.info(f"Using the grid size {_GRID_SIZE}")
+        initGrid(grid_size if grid_size is not None else [1, 1, 1, 1])
         if use_default_grid and not use_default_latt:
             _MPI_LOGGER.info(f"Using the lattice size {latt_size} only for getting the default grid size {_GRID_SIZE}")
         if use_default_latt:
             _DEFAULT_LATTICE = LatticeInfo(latt_size, t_boundary, anisotropy)
             _MPI_LOGGER.info(f"Using the default lattice LatticeInfo({latt_size}, {t_boundary}, {anisotropy})")
 
-        _initEnvironWarn(resource_path=resource_path if resource_path != "" else None)
-        _initEnviron(
+        _setEnvironWarn(resource_path=resource_path if resource_path != "" else None)
+        _setEnviron(
             rank_verbosity=",".join(rank_verbosity) if rank_verbosity != [0] else None,
             enable_mps="1" if enable_mps else None,
             enable_gdr="1" if enable_gdr else None,

pyquda_core/pyquda/_version.py

@@ -1 +1 @@
-__version__ = "0.9.10"
+__version__ = "0.9.11"

pyquda_core/pyquda/field.py

@@ -130,40 +130,47 @@ def lexico(data: numpy.ndarray, axes: List[int], dtype=None):
     Npre = int(numpy.prod(shape[: axes[0]]))
     Nsuf = int(numpy.prod(shape[axes[-1] + 1 :]))
     dtype = data.dtype if dtype is None else dtype
-    data_cb2 = data.reshape(Npre, 2, Lt, Lz, Ly, Lx // 2, Nsuf)
+    data_evenodd = data.reshape(Npre, 2, Lt, Lz, Ly, Lx // 2, Nsuf)
     data_lexico = numpy.zeros((Npre, Lt, Lz, Ly, Lx, Nsuf), dtype)
     for t in range(Lt):
         for z in range(Lz):
             for y in range(Ly):
                 eo = (t + z + y) % 2
                 if eo == 0:
-                    data_lexico[:, t, z, y, 0::2] = data_cb2[:, 0, t, z, y, :]
-                    data_lexico[:, t, z, y, 1::2] = data_cb2[:, 1, t, z, y, :]
+                    data_lexico[:, t, z, y, 0::2] = data_evenodd[:, 0, t, z, y, :]
+                    data_lexico[:, t, z, y, 1::2] = data_evenodd[:, 1, t, z, y, :]
                 else:
-                    data_lexico[:, t, z, y, 1::2] = data_cb2[:, 0, t, z, y, :]
-                    data_lexico[:, t, z, y, 0::2] = data_cb2[:, 1, t, z, y, :]
+                    data_lexico[:, t, z, y, 1::2] = data_evenodd[:, 0, t, z, y, :]
+                    data_lexico[:, t, z, y, 0::2] = data_evenodd[:, 1, t, z, y, :]
     return data_lexico.reshape(*shape[: axes[0]], Lt, Lz, Ly, Lx, *shape[axes[-1] + 1 :])
 
 
-def cb2(data: numpy.ndarray, axes: List[int], dtype=None):
+def evenodd(data: numpy.ndarray, axes: List[int], dtype=None):
     shape = data.shape
     Lt, Lz, Ly, Lx = [shape[axis] for axis in axes]
     Npre = int(numpy.prod(shape[: axes[0]]))
     Nsuf = int(numpy.prod(shape[axes[-1] + 1 :]))
     dtype = data.dtype if dtype is None else dtype
     data_lexico = data.reshape(Npre, Lt, Lz, Ly, Lx, Nsuf)
-    data_cb2 = numpy.zeros((Npre, 2, Lt, Lz, Ly, Lx // 2, Nsuf), dtype)
+    data_evenodd = numpy.zeros((Npre, 2, Lt, Lz, Ly, Lx // 2, Nsuf), dtype)
     for t in range(Lt):
         for z in range(Lz):
             for y in range(Ly):
                 eo = (t + z + y) % 2
                 if eo == 0:
-                    data_cb2[:, 0, t, z, y, :] = data_lexico[:, t, z, y, 0::2]
-                    data_cb2[:, 1, t, z, y, :] = data_lexico[:, t, z, y, 1::2]
+                    data_evenodd[:, 0, t, z, y, :] = data_lexico[:, t, z, y, 0::2]
+                    data_evenodd[:, 1, t, z, y, :] = data_lexico[:, t, z, y, 1::2]
                 else:
-                    data_cb2[:, 0, t, z, y, :] = data_lexico[:, t, z, y, 1::2]
-                    data_cb2[:, 1, t, z, y, :] = data_lexico[:, t, z, y, 0::2]
-    return data_cb2.reshape(*shape[: axes[0]], 2, Lt, Lz, Ly, Lx // 2, *shape[axes[-1] + 1 :])
+                    data_evenodd[:, 0, t, z, y, :] = data_lexico[:, t, z, y, 1::2]
+                    data_evenodd[:, 1, t, z, y, :] = data_lexico[:, t, z, y, 0::2]
+    return data_evenodd.reshape(*shape[: axes[0]], 2, Lt, Lz, Ly, Lx // 2, *shape[axes[-1] + 1 :])
+
+
+def cb2(data: numpy.ndarray, axes: List[int], dtype=None):
+    from . import getLogger
+
+    getLogger().warning("cb2 is deprecated, use evenodd instead", DeprecationWarning)
+    return evenodd(data, axes, dtype)
 
 
 def checksum(latt_info: Union[LatticeInfo, GeneralInfo], data: numpy.ndarray) -> Tuple[int, int]:
@@ -675,9 +682,9 @@ def load(
         if Nc is not None:
             latt_info.Nc = Nc
         if not issubclass(cls, MultiField):
-            retval = cls(latt_info, cb2(value, [0, 1, 2, 3]))
+            retval = cls(latt_info, evenodd(value, [0, 1, 2, 3]))
         else:
-            retval = cls(latt_info, len(label), numpy.asarray([cb2(data, [0, 1, 2, 3]) for data in value]))
+            retval = cls(latt_info, len(label), numpy.asarray([evenodd(data, [0, 1, 2, 3]) for data in value]))
         secs = perf_counter() - s
         getLogger().debug(f"Loaded {filename} in {secs:.3f} secs, {gbytes / secs:.3f} GB/s")
         return retval

pyquda_utils/core.py

@@ -4,6 +4,7 @@
 
 from pyquda import (
     initGPU,
+    initGrid,
     initQUDA,
     init,
     getCoordFromRank,
@@ -15,6 +16,7 @@
     getGridCoord,
     setDefaultLattice,
     getDefaultLattice,
+    getCUDABackend,
     getLogger,
     setLoggerLevel,
     dirac as fermion,
@@ -41,7 +43,8 @@
     LatticePropagator,
     LatticeStaggeredPropagator,
     lexico,
-    cb2,
+    evenodd,
+    evenodd as cb2,
 )
 from pyquda.dirac.abstract import Multigrid, FermionDirac, StaggeredFermionDirac
 

pyquda_utils/deprecated.py

@@ -1,6 +1,6 @@
 from typing import List
 
-from pyquda import getLogger, getGridSize, quda, enum_quda
+from pyquda import getLogger, getGridSize, pyquda as quda, enum_quda
 from pyquda.field import LatticeFermion, LatticeGauge, LatticeInfo, LatticePropagator, Nc, Ns
 from pyquda.dirac.abstract import FermionDirac
 
@@ -101,11 +101,11 @@ def getDslash(
     latt_info = LatticeInfo([Lx, Ly, Lz, Lt], t_boundary, xi)
 
     if clover_csw != 0.0:
-        from .dirac.clover_wilson import CloverWilsonDirac
+        from pyquda.dirac.clover_wilson import CloverWilsonDirac
 
         return CloverWilsonDirac(latt_info, mass, tol, maxiter, clover_csw, clover_xi, geo_block_size)
     else:
-        from .dirac.wilson import WilsonDirac
+        from pyquda.dirac.wilson import WilsonDirac
 
         return WilsonDirac(latt_info, mass, tol, maxiter, geo_block_size)
 
@@ -131,6 +131,6 @@ def getStaggeredDslash(
         t_boundary = 1
     latt_info = LatticeInfo([Lx, Ly, Lz, Lt], t_boundary, 1.0)
 
-    from .dirac.hisq import HISQDirac
+    from pyquda.dirac.hisq import HISQDirac
 
     return HISQDirac(latt_info, mass, tol, maxiter, naik_epsilon, None)

pyquda_utils/gpt.py

@@ -1,15 +1,17 @@
 from typing import List
 import numpy
 
-from pyquda import getSublatticeSize, getGridSize
-from pyquda.field import cb2, LatticeGauge, LatticeInfo, LatticePropagator
+from .core import evenodd, getGridSize, LatticeGauge, LatticeInfo, LatticePropagator
 
 import gpt as g
 
 
 def LatticeInfoGPT(grid: g.grid, gen_simd_width: int):
     assert getGridSize() == grid.mpi
-    sublatt_size = getSublatticeSize(grid.fdimensions, grid.mpi)
+    GLx, GLy, GLz, GLt = grid.fdimensions
+    Gx, Gy, Gz, Gt = grid.mpi
+    Lx, Ly, Lz, Lt = GLx // Gx, GLy // Gy, GLz // Gz, GLt // Gt
+    sublatt_size = [Lx, Ly, Lz, Lt]
     Nd = len(sublatt_size)
     precision = grid.precision.nbytes
     n_simd = gen_simd_width // (2 * precision)
@@ -32,7 +34,7 @@ def LatticeGaugeGPT(lattice: List[g.lattice], gen_simd_width: int, gauge: Lattic
         value = []
         for index in range(latt_info.Nd):
             value.append(
-                cb2(
+                evenodd(
                     numpy.asarray(lattice[index].mview()[0])
                     .view(f"<c{2 * gpt_prec}")
                     .reshape(*gpt_latt[::-1], Nc, Nc, *gpt_simd[::-1])
@@ -49,7 +51,8 @@ def LatticeGaugeGPT(lattice: List[g.lattice], gen_simd_width: int, gauge: Lattic
         for index in range(latt_info.Nd):
             gpt_shape = [i for sl in zip(gpt_simd, gpt_latt) for i in sl]
             lattice[index].mview()[0][:] = (
-                gauge[index].lexico()
+                gauge[index]
+                .lexico()
                 .astype(f"<c{2 * gpt_prec}")
                 .reshape(*gpt_shape, Nc, Nc)
                 .transpose(1, 3, 5, 7, 8, 9, 0, 2, 4, 6)
@@ -65,7 +68,7 @@ def LatticePropagatorGPT(lattice: g.lattice, gen_simd_width: int, propagator: La
     Ns, Nc = latt_info.Ns, latt_info.Nc
     assert lattice.describe().startswith(f"ot_matrix_spin_color({Ns},{Nc})")
     if propagator is None:
-        value = cb2(
+        value = evenodd(
             numpy.asarray(lattice.mview()[0])
             .view(f"<c{2 * gpt_prec}")
             .reshape(*gpt_latt[::-1], Ns, Ns, Nc, Nc, *gpt_simd[::-1])