def cublasXtDgemm(handle, transa, transb, m, n, k, alpha, A, lda, B, ldb,
beta, C, ldc):
status = _libcublas.cublasXtDgemm(handle, transa, transb, m, n, k,
ctypes.byref(ctypes.c_double(alpha)),
int(A), lda, int(B), ldb,
ctypes.byref(ctypes.c_double(beta)),
int(C), ldc)
cublasCheckStatus(status)
def cublasXtZgemm(handle, transa, transb, m, n, k, alpha, A, lda, B, ldb, beta,
C, ldc):
status = _libcublas.cublasXtZgemm(
handle, transa, transb, m, n, k,
ctypes.byref(cuda.cuDoubleComplex(alpha.real, alpha.imag)), int(A),
lda, int(B), ldb,
ctypes.byref(cuda.cuDoubleComplex(beta.real, beta.imag)), int(C), ldc)
cublasCheckStatus(status)
def cublasXtZgemm(handle, transa, transb, m, n, k, alpha, A, lda, B, ldb,
beta, C, ldc):
status = _libcublas.cublasXtZgemm(handle, transa, transb, m, n, k,
ctypes.byref(cuda.cuDoubleComplex(alpha.real,
alpha.imag)),
int(A), lda, int(B), ldb,
ctypes.byref(cuda.cuDoubleComplex(beta.real,
beta.imag)),
int(C), ldc)
cublasCheckStatus(status)
def cublasXtSetCpuRoutine(handle, blasOp, type, blasFunctor):
status = _libcublas.cublasXtSetCpuRoutine(handle, blasOp, type, blasFunctor)
cublasCheckStatus(status)
def cublasXtGetBlockDim(handle):
blockDim = ctypes.c_void_p()
status = _libcublas.cublasXtSetBlockDim(handle, ctypes.byref(blockDim))
cublasCheckStatus(status)
return blockDim.value
def cublasXtSetBlockDim(handle, blockDim):
status = _libcublas.cublasXtSetBlockDim(handle, blockDim)
cublasCheckStatus(status)
def cublasXtDeviceSelect(handle, nbDevices, deviceId):
status = _libcublas.cublasXtDeviceSelect(handle, nbDevices, deviceId)
cublasCheckStatus(status)
def cublasXtMaxBoards():
nbGpuBoards = ctypes.c_int()
status = _libcublas.cublasXtMaxBoards(ctypes.byref(nbGpuBoards))
cublasCheckStatus(status)
return nbGpuBoards.value
def cublasXtGetNumBoards(handle, deviceId):
nbBoards = ctypes.c_int()
status = _libcublas.cublasXtGetNumBoards(handle, deviceId, ctypes.byref(nbBoards))
cublasCheckStatus(status)
return nbBoards.value
def cublasXtDestroy(handle):
status = _libcublas.cublasXtDestroy(handle)
cublasCheckStatus(status)
def cublasXtCreate():
handle = ctypes.c_void_p()
status = _libcublas.cublasXtCreate(ctypes.byref(handle))
cublasCheckStatus(status)
return handle.value
def cublasXtSetCpuRatio(handle, blasOp, type, ratio):
status = _libcublas.cublasXtSetCpuRatio(handle, blasOp, type, ratio)
cublasCheckStatus(status)
def dot(x_gpu, y_gpu, transa='N', transb='N'):
"""
Dot product of two arrays.
For 1D arrays, this function computes the inner product. For 2D
arrays of shapes `(m, k)` and `(k, n)`, it computes the matrix
product; the result has shape `(m, n)`.
Parameters
----------
x_gpu : pycuda.gpuarray.GPUArray
Input array.
y_gpu : pycuda.gpuarray.GPUArray
Input array.
transa : char
If 'T', compute the product of the transpose of `x_gpu`.
If 'C', compute the product of the Hermitian of `x_gpu`.
transb : char
If 'T', compute the product of the transpose of `y_gpu`.
If 'C', compute the product of the Hermitian of `y_gpu`.
Returns
-------
c_gpu : pycuda.gpuarray.GPUArray, float{32,64}, or complex{64,128}
Inner product of `x_gpu` and `y_gpu`. When the inputs are 1D
arrays, the result will be returned as a scalar.
Notes
-----
The input matrices must all contain elements of the same data type.
Examples
--------
>>> import pycuda.gpuarray as gpuarray
>>> import pycuda.autoinit
>>> import numpy as np
>>> import linalg
>>> import misc
>>> linalg.init()
>>> a = np.asarray(np.random.rand(4, 2), np.float32)
>>> b = np.asarray(np.random.rand(2, 2), np.float32)
>>> a_gpu = gpuarray.to_gpu(a)
>>> b_gpu = gpuarray.to_gpu(b)
>>> c_gpu = linalg.dot(a_gpu, b_gpu)
>>> np.allclose(np.dot(a, b), c_gpu.get())
True
>>> d = np.asarray(np.random.rand(5), np.float32)
>>> e = np.asarray(np.random.rand(5), np.float32)
>>> d_gpu = gpuarray.to_gpu(d)
>>> e_gpu = gpuarray.to_gpu(e)
>>> f = linalg.dot(d_gpu, e_gpu)
>>> np.allclose(np.dot(d, e), f)
True
"""
if len(x_gpu.shape) == 1 and len(y_gpu.shape) == 1:
if x_gpu.size != y_gpu.size:
raise ValueError('arrays must be of same length')
# Compute inner product for 1D arrays:
if (x_gpu.dtype == np.complex64 and y_gpu.dtype == np.complex64):
cublas_func = cublas._libcublas.cublasCdotu
elif (x_gpu.dtype == np.float32 and y_gpu.dtype == np.float32):
cublas_func = cublas._libcublas.cublasSdot
elif (x_gpu.dtype == np.complex128 and y_gpu.dtype == np.complex128):
cublas_func = cublas._libcublas.cublasZdotu
elif (x_gpu.dtype == np.float64 and y_gpu.dtype == np.float64):
cublas_func = cublas._libcublas.cublasDdot
else:
raise ValueError('unsupported combination of input types')
result = cublas_func(x_gpu.size, int(x_gpu.gpudata), 1,
int(y_gpu.gpudata), 1)
if x_gpu.dtype == np.complex64:
return np.float32(result.x) + 1j * np.float32(result.y)
elif x_gpu.dtype == np.complex128:
return np.float64(result.x) + 1j * np.float64(result.y)
elif x_gpu.dtype == np.float32:
return np.float32(result)
else:
return np.float64(result)
else:
# Perform matrix multiplication for 2D arrays:
if (x_gpu.dtype == np.complex64 and y_gpu.dtype == np.complex64):
cublas_func = cublas._libcublas.cublasCgemm
alpha = cuda.cuFloatComplex(1, 0)
beta = cuda.cuFloatComplex(0, 0)
elif (x_gpu.dtype == np.float32 and y_gpu.dtype == np.float32):
cublas_func = cublas._libcublas.cublasSgemm
alpha = np.float32(1.0)
beta = np.float32(0.0)
elif (x_gpu.dtype == np.complex128 and y_gpu.dtype == np.complex128):
cublas_func = cublas._libcublas.cublasZgemm
alpha = cuda.cuDoubleComplex(1, 0)
beta = cuda.cuDoubleComplex(0, 0)
elif (x_gpu.dtype == np.float64 and y_gpu.dtype == np.float64):
cublas_func = cublas._libcublas.cublasDgemm
alpha = np.float64(1.0)
beta = np.float64(0.0)
else:
raise ValueError('unsupported combination of input types')
transa = lower(transa)
transb = lower(transb)
if transb in ['t', 'c']:
m, k = y_gpu.shape
elif transb in ['n']:
k, m = y_gpu.shape
else:
raise ValueError('invalid value for transb')
if transa in ['t', 'c']:
l, n = x_gpu.shape
elif transa in ['n']:
n, l = x_gpu.shape
else:
raise ValueError('invalid value for transa')
if l != k:
raise ValueError('objects are not aligned')
if transb == 'n':
lda = max(1, m)
else:
lda = max(1, k)
if transa == 'n':
ldb = max(1, k)
else:
ldb = max(1, n)
ldc = max(1, m)
# Note that the desired shape of the output matrix is the transpose
# of what CUBLAS assumes:
c_gpu = gpuarray.empty((n, ldc), x_gpu.dtype)
cublas_func(transb, transa, m, n, k, alpha, int(y_gpu.gpudata), lda,
int(x_gpu.gpudata), ldb, beta, int(c_gpu.gpudata), ldc)
status = cublas.cublasGetError()
cublas.cublasCheckStatus(status)
return c_gpu
def dot(x_gpu, y_gpu, transa='N', transb='N'):
"""
Dot product of two arrays.
For 1D arrays, this function computes the inner product. For 2D
arrays of shapes `(m, k)` and `(k, n)`, it computes the matrix
product; the result has shape `(m, n)`.
Parameters
----------
x_gpu : pycuda.gpuarray.GPUArray
Input array.
y_gpu : pycuda.gpuarray.GPUArray
Input array.
transa : char
If 'T', compute the product of the transpose of `x_gpu`.
If 'C', compute the product of the Hermitian of `x_gpu`.
transb : char
If 'T', compute the product of the transpose of `y_gpu`.
If 'C', compute the product of the Hermitian of `y_gpu`.
Returns
-------
c_gpu : pycuda.gpuarray.GPUArray, float{32,64}, or complex{64,128}
Inner product of `x_gpu` and `y_gpu`. When the inputs are 1D
arrays, the result will be returned as a scalar.
Notes
-----
The input matrices must all contain elements of the same data type.
Examples
--------
>>> import pycuda.gpuarray as gpuarray
>>> import pycuda.autoinit
>>> import numpy as np
>>> import linalg
>>> import misc
>>> linalg.init()
>>> a = np.asarray(np.random.rand(4, 2), np.float32)
>>> b = np.asarray(np.random.rand(2, 2), np.float32)
>>> a_gpu = gpuarray.to_gpu(a)
>>> b_gpu = gpuarray.to_gpu(b)
>>> c_gpu = linalg.dot(a_gpu, b_gpu)
>>> np.allclose(np.dot(a, b), c_gpu.get())
True
>>> d = np.asarray(np.random.rand(5), np.float32)
>>> e = np.asarray(np.random.rand(5), np.float32)
>>> d_gpu = gpuarray.to_gpu(d)
>>> e_gpu = gpuarray.to_gpu(e)
>>> f = linalg.dot(d_gpu, e_gpu)
>>> np.allclose(np.dot(d, e), f)
True
"""
if len(x_gpu.shape) == 1 and len(y_gpu.shape) == 1:
if x_gpu.size != y_gpu.size:
raise ValueError('arrays must be of same length')
# Compute inner product for 1D arrays:
if (x_gpu.dtype == np.complex64 and y_gpu.dtype == np.complex64):
cublas_func = cublas._libcublas.cublasCdotu
elif (x_gpu.dtype == np.float32 and y_gpu.dtype == np.float32):
cublas_func = cublas._libcublas.cublasSdot
elif (x_gpu.dtype == np.complex128 and y_gpu.dtype == np.complex128):
cublas_func = cublas._libcublas.cublasZdotu
elif (x_gpu.dtype == np.float64 and y_gpu.dtype == np.float64):
cublas_func = cublas._libcublas.cublasDdot
else:
raise ValueError('unsupported combination of input types')
result = cublas_func(x_gpu.size, int(x_gpu.gpudata), 1,
int(y_gpu.gpudata), 1)
if x_gpu.dtype == np.complex64:
return np.float32(result.x)+1j*np.float32(result.y)
elif x_gpu.dtype == np.complex128:
return np.float64(result.x)+1j*np.float64(result.y)
elif x_gpu.dtype == np.float32:
return np.float32(result)
else:
return np.float64(result)
else:
# Perform matrix multiplication for 2D arrays:
if (x_gpu.dtype == np.complex64 and y_gpu.dtype == np.complex64):
cublas_func = cublas._libcublas.cublasCgemm
alpha = cuda.cuFloatComplex(1, 0)
beta = cuda.cuFloatComplex(0, 0)
elif (x_gpu.dtype == np.float32 and y_gpu.dtype == np.float32):
cublas_func = cublas._libcublas.cublasSgemm
alpha = np.float32(1.0)
beta = np.float32(0.0)
elif (x_gpu.dtype == np.complex128 and y_gpu.dtype == np.complex128):
cublas_func = cublas._libcublas.cublasZgemm
alpha = cuda.cuDoubleComplex(1, 0)
beta = cuda.cuDoubleComplex(0, 0)
elif (x_gpu.dtype == np.float64 and y_gpu.dtype == np.float64):
cublas_func = cublas._libcublas.cublasDgemm
alpha = np.float64(1.0)
beta = np.float64(0.0)
else:
raise ValueError('unsupported combination of input types')
transa = lower(transa)
transb = lower(transb)
if transb in ['t', 'c']:
m, k = y_gpu.shape
elif transb in ['n']:
k, m = y_gpu.shape
else:
raise ValueError('invalid value for transb')
if transa in ['t', 'c']:
l, n = x_gpu.shape
elif transa in ['n']:
n, l = x_gpu.shape
else:
raise ValueError('invalid value for transa')
if l != k:
raise ValueError('objects are not aligned')
if transb == 'n':
lda = max(1, m)
else:
lda = max(1, k)
if transa == 'n':
ldb = max(1, k)
else:
ldb = max(1, n)
ldc = max(1, m)
# Note that the desired shape of the output matrix is the transpose
# of what CUBLAS assumes:
c_gpu = gpuarray.empty((n, ldc), x_gpu.dtype)
cublas_func(transb, transa, m, n, k, alpha, int(y_gpu.gpudata),
lda, int(x_gpu.gpudata), ldb, beta, int(c_gpu.gpudata), ldc)
status = cublas.cublasGetError()
cublas.cublasCheckStatus(status)
return c_gpu
def dot(a_gpu, b_gpu):
"""
Matrix product of two arrays.
For 1D arrays, this function computes the inner product. For 2D
arrays of shapes `(m, k)` and `(k, n)`, it computes the matrix
product; the result has shape `(m, n)`.
Parameters
----------
a_gpu : pycuda.gpuarray.GPUArray
Input array.
b_gpu : pycuda.gpuarray.GPUArray
Input array.
Returns
-------
c_gpu : pycuda.gpuarray.GPUArray
Dot product of `a_gpu` and `b_gpu`.
Notes
-----
The input matrices must all contain elements of the same data type.
Examples
--------
>>> import pycuda.gpuarray as gpuarray
>>> import pycuda.autoinit
>>> import numpy as np
>>> import linalg
>>> linalg.init()
>>> a = np.asarray(np.random.rand(4, 2), np.float32)
>>> b = np.asarray(np.random.rand(2, 2), np.float32)
>>> a_gpu = gpuarray.to_gpu(a)
>>> b_gpu = gpuarray.to_gpu(b)
>>> c_gpu = linalg.dot(a_gpu, b_gpu)
>>> np.allclose(np.dot(a, b), c_gpu.get())
True
>>> d = np.asarray(np.random.rand(5), np.float32)
>>> e = np.asarray(np.random.rand(5), np.float32)
>>> d_gpu = gpuarray.to_gpu(d)
>>> e_gpu = gpuarray.to_gpu(e)
>>> f = linalg.dot(d_gpu, e_gpu)
>>> np.allclose(np.dot(d, e), f)
True
>>> p = np.asarray(np.random.rand(4, 2), np.complex64)
>>> q = np.asarray(np.random.rand(2, 2), np.complex64)
>>> p_gpu = gpuarray.to_gpu(p)
>>> q_gpu = gpuarray.to_gpu(q)
>>> r_gpu = linalg.dot(p_gpu, q_gpu)
>>> np.allclose(np.dot(p, q), r_gpu.get())
True
>>> s = np.asarray(np.random.rand(5), np.complex128)
>>> t = np.asarray(np.random.rand(5), np.complex128)
>>> s_gpu = gpuarray.to_gpu(s)
>>> t_gpu = gpuarray.to_gpu(t)
>>> u = linalg.dot(s_gpu, t_gpu)
>>> np.allclose(np.dot(s, t), u)
True
"""
if len(a_gpu.shape) == 1 and len(b_gpu.shape) == 1:
# Compute inner product for 1D arrays:
if (a_gpu.dtype == np.complex64 and b_gpu.dtype == np.complex64):
cublas_func = cublas._libcublas.cublasCdotu
elif (a_gpu.dtype == np.float32 and b_gpu.dtype == np.float32):
cublas_func = cublas._libcublas.cublasSdot
elif (a_gpu.dtype == np.complex128 and b_gpu.dtype == np.complex128):
cublas_func = cublas._libcublas.cublasZdotu
elif (a_gpu.dtype == np.float64 and b_gpu.dtype == np.float64):
cublas_func = cublas._libcublas.cublasDdot
else:
raise ValueError('unsupported combination of input types')
result = cublas_func(a_gpu.size, int(a_gpu.gpudata), 1,
int(b_gpu.gpudata), 1)
if a_gpu.dtype == np.complex64:
return np.float32(result.x)+1j*np.float32(result.y)
elif a_gpu.dtype == np.complex128:
return np.float64(result.x)+1j*np.float64(result.y)
elif a_gpu.dtype == np.float32:
return np.float32(result)
else:
return np.float64(result)
else:
# Perform matrix multiplication for 2D arrays:
if (a_gpu.dtype == np.complex64 and b_gpu.dtype == np.complex64):
cublas_func = cublas._libcublas.cublasCgemm
alpha = cuda.cuFloatComplex(1, 0)
beta = cuda.cuFloatComplex(0, 0)
elif (a_gpu.dtype == np.float32 and b_gpu.dtype == np.float32):
cublas_func = cublas._libcublas.cublasSgemm
alpha = np.float32(1.0)
beta = np.float32(0.0)
elif (a_gpu.dtype == np.complex128 and b_gpu.dtype == np.complex128):
cublas_func = cublas._libcublas.cublasZgemm
alpha = cuda.cuDoubleComplex(1, 0)
beta = cuda.cuDoubleComplex(0, 0)
elif (a_gpu.dtype == np.float64 and b_gpu.dtype == np.float64):
cublas_func = cublas._libcublas.cublasDgemm
alpha = np.float64(1.0)
beta = np.float64(0.0)
else:
raise ValueError('unsupported combination of input types')
transa = 'N'
transb = 'N'
m = b_gpu.shape[1]
n = a_gpu.shape[0]
k = b_gpu.shape[0]
lda = m
ldb = k
ldc = max(1, m)
c_gpu = gpuarray.empty((a_gpu.shape[0], b_gpu.shape[1]), a_gpu.dtype)
cublas_func(transb, transa, m, n, k, alpha, int(b_gpu.gpudata),
lda, int(a_gpu.gpudata), ldb, beta, int(c_gpu.gpudata), ldc)
status = cublas.cublasGetError()
cublas.cublasCheckStatus(status)
return c_gpu