def rk(alpha, a, beta, c, trans='c'):
"""Rank-k update of a matrix."""
assert isinstance(a, mic.OffloadArray)
assert isinstance(c, mic.OffloadArray)
dt = map_dtype(a.dtype)
# determine sizes of the matrices
am = a.shape[0]
ak = np.prod(a.shape[1:])
ck = c.shape[0]
cn = np.prod(c.shape[1:])
n, k = am, ak
ldc = c.array.strides[0] / c.array.strides[1]
if a.dtype in [np.complex]:
alpha = complex(alpha)
beta = complex(beta)
# perform the offload
stream.invoke(library.mic_syrk, dt, a, c, n, k,
ldc, alpha, beta)
stream.sync()
def gemm(alpha, a, b, beta, c, transa='n'):
# we want to make sure that we only use OffloadArrays here
assert isinstance(a, mic.OffloadArray)
assert isinstance(b, mic.OffloadArray)
assert isinstance(c, mic.OffloadArray)
# determine the datatype and map it to int
dt = map_dtype(a.dtype)
# determine sizes of the matrices
am = a.shape[0]
ak = np.prod(a.shape[1:])
bk = b.shape[0]
bn = np.prod(b.shape[1:])
cm = c.shape[0]
cn = np.prod(c.shape[1:])
# just some safety checks
if transa == 'n':
assert am == bn
assert ak == cn
assert bk == cm
trans = 0
m, n, k = ak, bk, am
lda = a.array.strides[0] / a.array.strides[-1]
ldb = b.array.strides[0] / b.array.strides[1]
ldc = c.array.strides[0] / c.array.strides[-1]
else:
assert am == cn
assert ak == bn
assert bk == cm
trans = 1
m, n, k = am, bk, ak
lda = k
ldb = b.array.strides[0] / b.array.strides[-1]
ldc = c.array.strides[0] / c.array.strides[1]
if a.dtype in [np.complex]:
alpha = complex(alpha)
beta = complex(beta)
# perform the offload
stream.invoke(library.mic_gemm, dt,
a, b, c,
m, n, k,
lda, ldb, ldc,
alpha, beta, trans)
stream.sync()
def gemm(alpha, a, b, beta, c, transa='n'):
# we want to make sure that we only use OffloadArrays here
assert isinstance(a, mic.OffloadArray)
assert isinstance(b, mic.OffloadArray)
assert isinstance(c, mic.OffloadArray)
# determine the datatype and map it to int
dt = map_dtype(a.dtype)
# determine sizes of the matrices
am = a.shape[0]
ak = np.prod(a.shape[1:])
bk = b.shape[0]
bn = np.prod(b.shape[1:])
cm = c.shape[0]
cn = np.prod(c.shape[1:])
# just some safety checks
if transa == 'n':
assert am == bn
assert ak == cn
assert bk == cm
trans = 0
m, n, k = ak, bk, am
lda = a.array.strides[0] / a.array.strides[-1]
ldb = b.array.strides[0] / b.array.strides[1]
ldc = c.array.strides[0] / c.array.strides[-1]
else:
assert am == cn
assert ak == bn
assert bk == cm
trans = 1
m, n, k = am, bk, ak
lda = k
ldb = b.array.strides[0] / b.array.strides[-1]
ldc = c.array.strides[0] / c.array.strides[1]
if a.dtype in [np.complex]:
alpha = complex(alpha)
beta = complex(beta)
# perform the offload
stream.invoke(library.mic_gemm, dt, a, b, c, m, n, k, lda, ldb, ldc, alpha,
beta, trans)
stream.sync()
def r2k(alpha, a, b, beta, c):
"""Rank-2k update of a matrix."""
assert isinstance(a, mic.OffloadArray)
assert isinstance(b, mic.OffloadArray)
assert isinstance(c, mic.OffloadArray)
assert (map_dtype(a.dtype) != 2)
# determine sizes of the matrices
am = a.shape[0]
ak = np.prod(a.shape[1:])
bm = b.shape[0]
bk = np.prod(b.shape[1:])
ck = c.shape[0]
cn = np.prod(c.shape[1:])
n, k = am, ak
ldc = c.array.strides[0] / c.array.strides[1]
stream.invoke(library.mic_dsyr2k, a, b, c, n, k, ldc, alpha, beta)
stream.sync()
def r2k(alpha, a, b, beta, c):
"""Rank-2k update of a matrix."""
assert isinstance(a, mic.OffloadArray)
assert isinstance(b, mic.OffloadArray)
assert isinstance(c, mic.OffloadArray)
assert(map_dtype(a.dtype) != 2)
# determine sizes of the matrices
am = a.shape[0]
ak = np.prod(a.shape[1:])
bm = b.shape[0]
bk = np.prod(b.shape[1:])
ck = c.shape[0]
cn = np.prod(c.shape[1:])
n, k = am, ak
ldc = c.array.strides[0] / c.array.strides[1]
stream.invoke(library.mic_dsyr2k, a, b, c, n, k,
ldc, alpha, beta)
stream.sync()
def rk(alpha, a, beta, c, trans='c'):
"""Rank-k update of a matrix."""
assert isinstance(a, mic.OffloadArray)
assert isinstance(c, mic.OffloadArray)
dt = map_dtype(a.dtype)
# determine sizes of the matrices
am = a.shape[0]
ak = np.prod(a.shape[1:])
ck = c.shape[0]
cn = np.prod(c.shape[1:])
n, k = am, ak
ldc = c.array.strides[0] / c.array.strides[1]
if a.dtype in [np.complex]:
alpha = complex(alpha)
beta = complex(beta)
# perform the offload
stream.invoke(library.mic_syrk, dt, a, c, n, k, ldc, alpha, beta)
stream.sync()
