Skip to content
This repository has been archived by the owner on Jan 3, 2023. It is now read-only.

add openblas support in FindCBLAS.cmake #1

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

add openblas support in FindCBLAS.cmake #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
56dde5a
add openblas support in FindCBLAS.cmake
May 9, 2015
7001be9
fix bug caused by improper use of find_library() in ctypes
May 9, 2015
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 13 additions & 0 deletions cmake/FindCBLAS.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -159,6 +159,19 @@ IF(NOT CBLAS_LIBRARIES)
)
ENDIF(NOT CBLAS_LIBRARIES)

IF(NOT CBLAS_LIBRARIES)
# CBLAS in OpenBLAS library?
CHECK_ALL_LIBRARIES(
CBLAS_LIBRARIES
CBLAS
cblas_dgemm
""
"openblas"
"cblas.h"
TRUE
)
ENDIF(NOT CBLAS_LIBRARIES)

IF(NOT CBLAS_LIBRARIES)
# CBLAS in ATLAS library? (http://math-atlas.sourceforge.net/)
CHECK_ALL_LIBRARIES(
Expand Down
87 changes: 42 additions & 45 deletions cudanet/cudanet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,10 +10,7 @@

MAX_ONES = 1024*256

cudanet_lib_path = ct.util.find_library('cconv2_cudanet')
if cudanet_lib_path is None:
raise OSError("Problems locating libcudanet shared library")
_cudanet = ct.cdll.LoadLibrary(cudanet_lib_path)
_cudanet = ct.cdll.LoadLibrary("libcconv2_cudanet.so")


_cudanet.get_last_cuda_error.restype = ct.c_char_p
Expand Down Expand Up @@ -255,7 +252,7 @@ class cudanetmat(ct.Structure):
('owns_data', ct.c_int)]

class rnd_struct(ct.Structure):
_fields_ = [('dev_rnd_mults', ct.POINTER(ct.c_uint)),
_fields_ = [('dev_rnd_mults', ct.POINTER(ct.c_uint)),
('dev_rnd_words', ct.POINTER(ct.c_longlong))]


Expand Down Expand Up @@ -479,7 +476,7 @@ def assign(self, val):
err_code = _cudanet.assign_scalar(self.p_mat, ct.c_float(val))
else:
raise ValueError("Assigned value must be of type CUDAMatrix, int, or float.")

if err_code:
raise generate_exception(err_code)

Expand Down Expand Up @@ -540,7 +537,7 @@ def subtract_dot(self, m1, m2, mult = 1., beta = 1.):
Subtract the dot product of m1 and m2 from the matrix, scaled by mult.
Self is scaled by beta before subtracting anything.
"""

return self.add_dot(m1, m2, mult = -1. * mult, beta = beta)
def add_mult(self, mat2, alpha = 1.):
"""
Expand All @@ -552,7 +549,7 @@ def add_mult(self, mat2, alpha = 1.):
raise generate_exception(err_code)

return self

def subtract_mult(self, mat2, alpha = 1.):
"""
Subtract a multiple of mat2 from the matrix.
Expand Down Expand Up @@ -730,7 +727,7 @@ def slice(self, first_col, last_col, include_host = False):
except:
new_mat.sliceof = self

# reproduce the slice on the host as well (if requested)
# reproduce the slice on the host as well (if requested)
if include_host and self.mat.on_host:
new_mat.numpy_array = self.numpy_array[:, first_col:last_col]
_cudanet.set_host_mat(new_mat.p_mat, new_mat.numpy_array.ctypes.data_as(ct.POINTER(ct.c_float)))
Expand Down Expand Up @@ -841,7 +838,7 @@ def sum(self, axis, target = None, mult = 1.):
if axis == 0:
if not target:
target = empty((1, n))

elif axis == 1:
if not target:
target = empty((m, 1))
Expand Down Expand Up @@ -1010,7 +1007,7 @@ def min(self, axis, target = None):
if axis == 0:
if not target:
target = empty((1, n))

elif axis == 1:
if not target:
target = empty((m, 1))
Expand All @@ -1037,7 +1034,7 @@ def max(self, axis, target = None):
if axis == 0:
if not target:
target = empty((1, n))

elif axis == 1:
if not target:
target = empty((m, 1))
Expand All @@ -1064,7 +1061,7 @@ def sum(self, axis, target = None):
if axis == 0:
if not target:
target = empty((1, n))

elif axis == 1:
if not target:
target = empty((m, 1))
Expand All @@ -1088,7 +1085,7 @@ def mean_norm(self, axis, target = None):

m, n = self.shape

if not target:
if not target:
target = empty((m,n))

err_code = _cudanet.mean_norm(self.p_mat, target.p_mat, ct.c_int(axis))
Expand Down Expand Up @@ -1133,7 +1130,7 @@ def argmax(self, axis, target = None):
if axis == 0:
if not target:
target = empty((1, n))

elif axis == 1:
if not target:
target = empty((m, 1))
Expand All @@ -1157,7 +1154,7 @@ def argmin(self, axis, target = None):
if axis == 0:
if not target:
target = empty((1, n))

elif axis == 1:
if not target:
target = empty((m, 1))
Expand Down Expand Up @@ -1234,7 +1231,7 @@ def print_devmat(self):
"""

_cudanet.print_devmat(self.p_mat)

def empty(shape):
"""
Creates and returns a new CUDAMatrix with the given shape.
Expand Down Expand Up @@ -1530,14 +1527,14 @@ def where(condition_mat, if_mat, else_mat, target = None):
def max_pool(imgs, target, channels, sizeX, paddingStart, moduleStride, numModulesX):
"""
Perform Max Pooling of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
"""

err_code = _cudanet.max_pool(imgs.p_mat, target.p_mat, ct.c_int(channels),
ct.c_int(sizeX), ct.c_int(paddingStart),
ct.c_int(moduleStride), ct.c_int(numModulesX))
ct.c_int(moduleStride), ct.c_int(numModulesX))
if err_code:
raise generate_exception(err_code)

Expand All @@ -1546,7 +1543,7 @@ def max_pool(imgs, target, channels, sizeX, paddingStart, moduleStride, numModul
def max_pool_undo(imgs, maxGrads, maxActs, target, sizeX, paddingStart, moduleStride, numModulesX):
"""
Undo Max Pooling of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride
"""
Expand All @@ -1555,7 +1552,7 @@ def max_pool_undo(imgs, maxGrads, maxActs, target, sizeX, paddingStart, moduleSt
target.p_mat, ct.c_int(sizeX),
ct.c_int(paddingStart),
ct.c_int(moduleStride),
ct.c_int(numModulesX))
ct.c_int(numModulesX))
if err_code:
raise generate_exception(err_code)

Expand All @@ -1564,14 +1561,14 @@ def max_pool_undo(imgs, maxGrads, maxActs, target, sizeX, paddingStart, moduleSt
def l2_pool(imgs, target, channels, sizeX, paddingStart, moduleStride, numModulesX):
"""
Perform L2 Pooling of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
"""

err_code = _cudanet.l2_pool(imgs.p_mat, target.p_mat, ct.c_int(channels),
ct.c_int(sizeX), ct.c_int(paddingStart),
ct.c_int(moduleStride), ct.c_int(numModulesX))
ct.c_int(moduleStride), ct.c_int(numModulesX))
if err_code:
raise generate_exception(err_code)

Expand All @@ -1580,7 +1577,7 @@ def l2_pool(imgs, target, channels, sizeX, paddingStart, moduleStride, numModule
def l2_pool_undo(imgs, l2Grads, l2Acts, target, sizeX, paddingStart, moduleStride, numModulesX):
"""
Undo L2 Pooling of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride
"""
Expand All @@ -1589,7 +1586,7 @@ def l2_pool_undo(imgs, l2Grads, l2Acts, target, sizeX, paddingStart, moduleStrid
target.p_mat, ct.c_int(sizeX),
ct.c_int(paddingStart),
ct.c_int(moduleStride),
ct.c_int(numModulesX))
ct.c_int(numModulesX))
if err_code:
raise generate_exception(err_code)

Expand All @@ -1598,14 +1595,14 @@ def l2_pool_undo(imgs, l2Grads, l2Acts, target, sizeX, paddingStart, moduleStrid
def avg_pool(imgs, target, channels, sizeX, paddingStart, moduleStride, numModulesX):
"""
Perform Max Pooling of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
"""

err_code = _cudanet.avg_pool(imgs.p_mat, target.p_mat, ct.c_int(channels),
ct.c_int(sizeX), ct.c_int(paddingStart),
ct.c_int(moduleStride), ct.c_int(numModulesX))
ct.c_int(moduleStride), ct.c_int(numModulesX))
if err_code:
raise generate_exception(err_code)

Expand All @@ -1614,7 +1611,7 @@ def avg_pool(imgs, target, channels, sizeX, paddingStart, moduleStride, numModul
def avg_pool_undo(avgGrads, target, sizeX, paddingStart, moduleStride, numModulesX, imgSizeX):
"""
Undo Avg Pooling of kernel dimension sizeX on imgs and put result in target
average Gradients as (CxHxW) Rows x (N) Columns in 'C' order
average Gradients as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride, int numImgColors, int numGroups
"""
Expand All @@ -1632,7 +1629,7 @@ def avg_pool_undo(avgGrads, target, sizeX, paddingStart, moduleStride, numModule
def unpool_forward(smallMat, largeMat, channels, sizeX, smallX, largeX):
"""
Undo Avg Pooling of kernel dimension sizeX on imgs and put result in target
average Gradients as (CxHxW) Rows x (N) Columns in 'C' order
average Gradients as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride, int numImgColors, int numGroups
"""
Expand All @@ -1648,7 +1645,7 @@ def unpool_forward(smallMat, largeMat, channels, sizeX, smallX, largeX):
def unpool_backward(largeMat, smallMat, channels, sizeX, smallX, largeX):
"""
Undo Avg Pooling of kernel dimension sizeX on imgs and put result in target
average Gradients as (CxHxW) Rows x (N) Columns in 'C' order
average Gradients as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride, int numImgColors, int numGroups
"""
Expand All @@ -1665,7 +1662,7 @@ def crossmap_response_norm(imgs, target, channels, sizeX, scale, power):
"""
Perform response normalization across channels of kernel dimension sizeX on
imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int channels, sizeX, float scale, power
"""
Expand All @@ -1683,7 +1680,7 @@ def crossmap_response_norm(imgs, target, channels, sizeX, scale, power):
def crossmap_response_norm_undo(imgs, respGrads, respActs, target, channels, sizeX, scale, power):
"""
Undo response normalization of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int channels, sizeX, float scale, power
"""
Expand All @@ -1704,7 +1701,7 @@ def local_contrast_norm(imgs, meanDiffs, denoms, target, imgSizeX, channels, siz
"""
Perform contrast normalization across channels of kernel dimension sizeX on
imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int channels, sizeX, float scale, power
"""
Expand All @@ -1725,11 +1722,11 @@ def local_contrast_norm(imgs, meanDiffs, denoms, target, imgSizeX, channels, siz
def local_contrast_norm_undo(meanDiffs, denoms, respGrads, respActs, target, channels, sizeX, scale, power):
"""
Undo contrast normalization of kernel dimension sizeX on imgs and put result in target
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int channels, sizeX, float scale, power
"""
err_code = _cudanet.local_contrast_norm_undo(meanDiffs.p_mat, denoms.p_mat,
err_code = _cudanet.local_contrast_norm_undo(meanDiffs.p_mat, denoms.p_mat,
respGrads.p_mat,
respActs.p_mat, target.p_mat,
ct.c_int(channels),
Expand Down Expand Up @@ -1766,7 +1763,7 @@ def convolution(wts, imgs, target, imgSizeY, numModulesY, numModulesX,
"""
Convolve wts with imgs and put result in target
Weights as (CxRxS) Rows x (K) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Images as (CxHxW) Rows x (N) Columns in 'C' order
Target as (KxPxQ) Rows x (N) Colums in 'C' order
int imgSizeY, int numModulesX, int paddingStart, int moduleStride,
int numImgColors, int numGroups
Expand All @@ -1793,9 +1790,9 @@ def deconvolve_errors(wts, errors, target, imgSizeY, imgSizeX, numModulesY,
"""
Backprop errors and put result in target
Weights as (CxRxS) Rows x (K) Columns in 'C' order
Errors as (KxPxQ) Rows x (N) Columns in 'C' order
Errors as (KxPxQ) Rows x (N) Columns in 'C' order
Target as (CxHxW) Rows x (N) Colums in 'C' order
int imgSizeY, ing imgSizeX, int numModulesY, int paddingStart,
int imgSizeY, ing imgSizeX, int numModulesY, int paddingStart,
int moduleStride, int numImgColors, int numGroups
"""
err_code = _cudanet.convolution_back_errors(wts.p_mat, errors.p_mat,
Expand All @@ -1818,7 +1815,7 @@ def deconvolve_wts(hidActs, imgs, target, imgSizeY, numModulesY, numModulesX,
numGroups, sumWidth, doLocal=False):
"""
Backprop acts grad with img grad to compute wts grad and put result in target
hidActs as (CxHxW) Rows x (N) Columns in 'C' order
hidActs as (CxHxW) Rows x (N) Columns in 'C' order
imgs as (KxPxQ) Rows x (N) Colums in 'C' order
Target as (CxRxS) Rows x (K) Columns in 'C' order
int imgSizeY, ing numModulesY, int numModulesX, int filterSize,
Expand Down Expand Up @@ -1863,7 +1860,7 @@ def xcov(X, Y, target = None, normX=1, normY=1, normAll=-1):
if (normX != 0 and normX != 1):
raise generate_exception(-6)
if (normY != 0 and normY != 1):
raise generate_exception(-6)
raise generate_exception(-6)

if (normAll == -1):
normFactor = np.float32(_cudanet.get_nonleading_dimension(X.p_mat))
Expand Down Expand Up @@ -1896,7 +1893,7 @@ def split(mat, nsplit, axis):
Meant to provide functionality similar to vsplit and split in numpy
Can split along either axis -- no default provided
Not streamed optimally at the moment, everything happens sequentially
each of the submats returned here have gpu buffers that are VIEWS of the
each of the submats returned here have gpu buffers that are VIEWS of the
original, they are not copied, and therefore don't "own" their data
"""
# Check validity of axis
Expand Down Expand Up @@ -1986,7 +1983,7 @@ def multi_way_error(probs, labels, labellogprob, top1probs, topkprobs, topk):
def softmax(mat, target = None, axis=0):
if not target:
target = empty(mat.shape)

err_code = _cudanet.softmax(mat.p_mat, target.p_mat, ct.c_int(axis))
if err_code:
raise generate_exception(err_code)
Expand Down