Skip to content

Commit

Permalink
AnalogCore Bugfixes for multi-param and GPU cores (#17)
Browse files Browse the repository at this point in the history
  • Loading branch information
@bfeinberg
bfeinberg authored Mar 25, 2024
1 parent fade4af commit b055e38
Show file tree
Hide file tree
Showing 2 changed files with 28 additions and 25 deletions.
2 changes: 1 addition & 1 deletion pyproject.toml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta"
[project]
name = "CrossSim"
description = "Accuracy simulator for analog in-memory computing"
version = "3.0"
version = "3.0.1"
license = {text = "BSD 3-Clause License"}
dependencies = [
"numpy",
Expand Down
51 changes: 27 additions & 24 deletions simulator/cores/analog_core.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -210,7 +210,10 @@ def __init__(
self.num_cores_row = self.Ncores // ((ncol - 1) // NrowsMax + 1)
self.num_cores_col = self.Ncores // self.num_cores_row
self.cores = [
[self._make_core(params_[r * c]) for c in range(self.num_cores_col)]
[
self._make_core(params_[r * self.num_cores_col + c])
for c in range(self.num_cores_col)
]
for r in range(self.num_cores_row)
]

Expand All @@ -224,9 +227,9 @@ def __init__(
# 3) Otherwise partition based on row/col_partition_strategy (max or even)

if nrow % self.num_cores_row == 0:
self.NrowsVec = (nrow // self.num_cores_row) * np.ones(
self.NrowsVec = (nrow // self.num_cores_row) * xp.ones(
self.num_cores_row,
dtype=np.int32,
dtype=xp.int32,
)
else:
# prio_partition = True in (
Expand Down Expand Up @@ -321,15 +324,15 @@ def set_matrix(
Raises:
ValueError: Matrix is not valid for the input parameters.
"""
matrix = np.asarray(matrix)
matrix = xp.asarray(matrix)
if self.shape != matrix.shape:
raise ValueError("Matrix shape must match AnalogCore shape")

if verbose:
print("Min/Max matrix values", np.min(matrix), np.max(matrix))
print("Min/Max matrix values", xp.min(matrix), xp.max(matrix))

if (
matrix.dtype == np.complex64 or matrix.dtype == np.complex128
matrix.dtype == xp.complex64 or matrix.dtype == xp.complex128
) and not self.complex_valued:
raise ValueError(
(
Expand All @@ -343,9 +346,9 @@ def set_matrix(
# Break up complex matrix into real and imaginary quadrants
if self.complex_valued:
Nx, Ny = matrix.shape
matrix_real = np.real(matrix)
matrix_imag = np.imag(matrix)
mcopy = np.zeros((2 * Nx, 2 * Ny), dtype=matrix_real.dtype)
matrix_real = xp.real(matrix)
matrix_imag = xp.imag(matrix)
mcopy = xp.zeros((2 * Nx, 2 * Ny), dtype=matrix_real.dtype)
mcopy[0:Nx, 0:Ny] = matrix_real
mcopy[Nx:, 0:Ny] = matrix_imag
mcopy[0:Nx, Ny:] = -matrix_imag
Expand All @@ -357,8 +360,8 @@ def set_matrix(
# If the values would exceed this range then you would have to reprogram all
# matrix values based on the new range, so instead we will clip and warn
if error_mask:
mat_max = np.max(matrix)
mat_min = np.min(matrix)
mat_max = xp.max(matrix)
mat_min = xp.min(matrix)

# Adding an epsilon here to avoid erroreous errors
if mat_max > (self.max + self._eps) or mat_min < (self.min - self._eps):
Expand Down Expand Up @@ -468,7 +471,7 @@ def get_matrix(self) -> npt.NDArray:
if self.Ncores == 1:
matrix = self.cores[0][0]._read_matrix()
else:
matrix = np.zeros((self.nrow, self.ncol))
matrix = xp.zeros((self.nrow, self.ncol))
for row, row_start, row_end in self.row_partition_bounds:
for col, col_start, col_end in self.col_partition_bounds:
matrix[row_start:row_end, col_start:col_end] = self.cores[row][
Expand Down Expand Up @@ -866,7 +869,7 @@ def __getitem__(self, item):
def __setitem__(self, key, value):
rslice, cslice, full_mask, _ = self._create_mask(key)
expanded_mat = self.get_matrix()
expanded_mat[rslice, cslice] = np.asarray(value)
expanded_mat[rslice, cslice] = xp.asarray(value)
error_mask = None if full_mask else (rslice, cslice)
self.set_matrix(expanded_mat, error_mask=error_mask)

Expand Down Expand Up @@ -950,21 +953,21 @@ def _set_limits_percentile(constraints, input_, reset=False):
"""
if (constraints.min is None or constraints.max is None) or reset:
if constraints.percentile >= 1.0:
X_max = np.max(np.abs(input_))
X_max = xp.max(xp.abs(input_))
X_max *= constraints.percentile
min_ = -X_max
max_ = X_max

elif constraints.percentile < 1.0:
X_posmax = np.percentile(input_, 100 * constraints.percentile)
X_negmax = np.percentile(input_, 100 - 100 * constraints.percentile)
X_max = np.max(np.abs(np.array([X_posmax, X_negmax])))
X_posmax = xp.percentile(input_, 100 * constraints.percentile)
X_negmax = xp.percentile(input_, 100 - 100 * constraints.percentile)
X_max = xp.max(xp.abs(xp.array([X_posmax, X_negmax])))
min_ = -X_max
max_ = X_max

# Ensure min_ and max_ aren't the same for uniform inputs
if min_ == max_:
eps = np.finfo(float).eps
eps = xp.finfo(float).eps
min_ -= eps
max_ += eps
return (min_, max_)
Expand Down Expand Up @@ -1054,7 +1057,7 @@ def set_matrix(
verbose: bool = False,
error_mask: tuple[slice, slice] | None = None,
) -> None:
matrix = np.asarray(matrix)
matrix = xp.asarray(matrix)
self.parent.set_matrix(matrix.T, verbose=verbose, error_mask=error_mask)

def matvec(self, vec: npt.ArrayLike) -> npt.NDArray:
Expand Down Expand Up @@ -1157,7 +1160,7 @@ def get_matrix(self) -> npt.NDArray:
def set_matrix(self, matrix: npt.ArrayLike, verbose: bool = False, error_mask=None):
# TODO: Do we need to do anything with error_mask here?
expanded_mat = self.parent.get_matrix()
expanded_mat[self.rslice, self.cslice] = np.asarray(matrix)
expanded_mat[self.rslice, self.cslice] = xp.asarray(matrix)
self.parent.set_matrix(
expanded_mat,
verbose=verbose,
Expand All @@ -1172,7 +1175,7 @@ def matvec(self, other: npt.ArrayLike) -> npt.NDArray:
f"Dimension mismatch: {self.shape}, {other.shape}",
)

vec_in = np.zeros(self.parent.shape[1], dtype=other.dtype)
vec_in = xp.zeros(self.parent.shape[1], dtype=other.dtype)
vec_in[self.cslice] = other.flatten()

vec_out = self.parent.matvec(vec_in)
Expand All @@ -1195,7 +1198,7 @@ def matmat(self, other: npt.ArrayLike) -> npt.NDArray:
# out-of-slice rows
# For col slices we're just leaving empty entires in the input matrix
# corrosponding to missing rows
mat_in = np.zeros((self.parent.shape[1], other.shape[1]), dtype=other.dtype)
mat_in = xp.zeros((self.parent.shape[1], other.shape[1]), dtype=other.dtype)
for i in range(self.parent.shape[1])[self.cslice]:
mat_in[i] = other[(i - self.cslice.start) // self.cslice.step]
mat_out = self.parent.matmat(mat_in)
Expand All @@ -1209,7 +1212,7 @@ def vecmat(self, other: npt.ArrayLike) -> npt.NDArray:
f"Dimension mismatch: {other.shape}, {self.shape} ",
)

vec_in = np.zeros(self.parent.shape[0], dtype=other.dtype)
vec_in = xp.zeros(self.parent.shape[0], dtype=other.dtype)
vec_in[self.rslice] = other.flatten()

vec_out = self.parent.vecmat(vec_in)
Expand All @@ -1228,7 +1231,7 @@ def rmatmat(self, other: npt.ArrayLike) -> npt.NDArray:
f"Dimension mismatch: {other.shape}, {self.shape}",
)

mat_in = np.zeros((other.shape[0], self.parent.shape[0]), dtype=other.dtype)
mat_in = xp.zeros((other.shape[0], self.parent.shape[0]), dtype=other.dtype)
for i in range(self.parent.shape[0])[self.rslice]:
mat_in.T[i] = other.T[(i - self.rslice.start) // self.rslice.step]

Expand Down

0 comments on commit b055e38

Please sign in to comment.