def time_multi(N, nargs, niter=100):
map_exprs = ["%s*x%s[i]" % (i, i) for i in xrange(nargs)]
arguments = ",".join("__global float *x%s" % i for i in xrange(nargs))
k = OCLReductionKernel2(np.float32,
neutral="0",
reduce_expr="a+b",
map_exprs=map_exprs,
arguments=arguments)
ins = [
OCLArray.from_array(np.ones(N, np.float32))
for _ in xrange(len(map_exprs))
]
outs = [OCLArray.empty(1, np.float32) for _ in xrange(len(map_exprs))]
from time import time
t = time()
for _ in xrange(niter):
k(*ins, outs=outs)
get_device().queue.finish()
t = (time() - t) / niter
print "multi reduction: result =", [float(out.get()) for out in outs]
print "multi reduction:\t\t%.2f ms" % (1000 * t)
return t
def time_simple(N, nargs, niter=100):
from gputools import OCLReductionKernel
map_exprs = ["%s*x[i]" % i for i in xrange(nargs)]
ks = [
OCLReductionKernel(np.float32,
neutral="0",
reduce_expr="a+b",
map_expr="%s*x[i]" % i,
arguments="__global float *x")
for i in xrange(len(map_exprs))
]
ins = [
OCLArray.from_array(np.ones(N, np.float32))
for _ in xrange(len(map_exprs))
]
outs = [OCLArray.empty(1, np.float32) for _ in xrange(len(map_exprs))]
from time import time
t = time()
for _ in xrange(niter):
for k, inn, out in zip(ks, ins, outs):
k(inn, out=out)
get_device().queue.finish()
t = (time() - t) / niter
print "simple reduction: result =", [float(out.get()) for out in outs]
print "simple reduction:\t\t%.2f ms" % (1000 * t)
return t
def time_np(dshape, niter=3):
d = np.empty(dshape, np.complex64)
get_device().queue.finish()
t = time()
for _ in xrange(niter):
np.fft.fftn(d)
get_device().queue.finish()
t = (time()-t)/niter
print "CPU\t\t\t%s\t\t%.2f ms"%(dshape, 1000.*t)
def time_np(dshape, niter=3):
d = np.empty(dshape, np.complex64)
get_device().queue.finish()
t = time()
for _ in range(niter):
np.fft.fftn(d)
get_device().queue.finish()
t = (time() - t) / niter
print("CPU\t\t\t%s\t\t%.2f ms" % (dshape, 1000. * t))
return t
def time_gpu(dshape, niter=100, fast_math=False):
d_g = OCLArray.empty(dshape, np.complex64)
get_device().queue.finish()
plan = fft_plan(dshape, fast_math=fast_math)
t = time()
for _ in xrange(niter):
fft(d_g, inplace=True, plan=plan)
get_device().queue.finish()
t = (time()-t)/niter
print "GPU (fast_math = %s)\t%s\t\t%.2f ms"%(fast_math, dshape, 1000.*t)
def time_gpu(dshape, niter=100, fast_math=False):
d_g = OCLArray.empty(dshape, np.complex64)
get_device().queue.finish()
plan = fft_plan(dshape, fast_math=fast_math)
t = time()
for _ in range(niter):
fft(d_g, inplace=True, plan=plan)
get_device().queue.finish()
t = (time() - t) / niter
print("GPU (fast_math = %s)\t%s\t\t%.2f ms" % (fast_math, dshape, 1000. * t))
return t
def fft(arr_obj,res_g = None,
inplace = False,
inverse = False,
batch = 1,
plan = None,
fast_math = True,
normalize = True):
""" (inverse) fourier trafo of 1-3D arrays
creates a new plan or uses the given plan
the transformed arr_obj should be either a
- numpy array:
returns the fft as numpy array (inplace is ignored)
- OCLArray of type complex64:
writes transform into res_g if given, to arr_obj if inplace
or returns a new OCLArray with the transform otherwise
"""
if plan is None:
if batch==1:
plan = Plan(arr_obj.shape, queue = get_device().queue,
normalize = normalize,
fast_math = fast_math)
else:
plan = Plan(arr_obj.shape[1:], queue = get_device().queue,
normalize = normalize,
fast_math = fast_math)
if isinstance(arr_obj,np.ndarray):
return _ocl_fft_numpy(plan, arr_obj,inverse = inverse)
elif isinstance(arr_obj,OCLArray):
if not arr_obj.dtype.type is np.complex64:
raise TypeError("OCLArray arr_obj has to be of complex64 type")
if inplace:
_ocl_fft_gpu_inplace(plan, arr_obj, inverse = inverse,
batch = batch)
else:
return _ocl_fft_gpu(plan, arr_obj, res_arr = res_g, inverse = inverse,
batch = batch)
else:
raise TypeError("array argument (1) has bad type: %s"%type(arr_obj))
def func_to_time(dshape, Niter = 10, **kwargs):
d = zeros(dshape,np.complex64)
d_g = gputools.OCLArray.from_array(d)
#burn in
f(d_g,**kwargs)
gputools.get_device().queue.finish()
t = time()
for _ in range(Niter):
f(d_g,**kwargs)
gputools.get_device().queue.finish()
return (time()-t)/Niter
def empty(cls, shape, dtype, num_channels=1, channel_order=None):
ctx = get_device().context
if not len(shape) in [2, 3]:
raise ValueError(
"dimension of shape wrong, should be 2...3 but is %s" %
len(shape))
mem_flags = cl.mem_flags.READ_WRITE
channel_type = cl.DTYPE_TO_CHANNEL_TYPE[np.dtype(dtype)]
_dict_channel_order = {
1: cl.channel_order.R,
2: cl.channel_order.RG,
3: cl.channel_order.RGB,
4: cl.channel_order.RGBA
}
if channel_order is None:
channel_order = _dict_channel_order[num_channels]
fmt = cl.ImageFormat(channel_order, channel_type)
res = cls(ctx, mem_flags, fmt, shape=shape[::-1])
res.dtype = dtype
res.num_channels = num_channels
return res
def write_array(self, data):
queue = get_device().queue
# 1d images dont have a shape but only a width
if hasattr(self, "shape"):
imshape = self.shape
else:
imshape = (self.width, )
ndim = len(imshape)
dshape = data.shape
# if clImg.format.channel_order in [cl.channel_order.RGBA,
# cl.channel_order.BGRA]:
# dshape = dshape[:-1]
if dshape != imshape[::-1]:
raise ValueError("write_array: wrong shape!", data.shape[::-1],
imshape)
else:
#cl.enqueue_write_image(queue,self,[0]*ndim,imshape,data)
#FIXME data.copy() is a work around
cl.enqueue_copy(queue,
self,
data.copy(),
origin=(0, ) * ndim,
region=imshape)
def estimate_niter(N):
"""returns niter s.t. the time spent on kernel is same as for memory transfer"""
a = np.ones(N,np.float32)
dev = get_device()
context, queue = dev.context, dev.queue
mf = cl.mem_flags
t = time()
copy_g = cl.Buffer(context, mf.ALLOC_HOST_PTR
,size = a.nbytes)
cl.enqueue_map_buffer(queue, copy_g, a,
device_offset=0,
is_blocking=False)
# cl.enqueue_copy(queue, copy_g, a,
# device_offset=0,
# is_blocking=False)
queue.flush()
# a_g = OCLArray.from_array(a, async = True)
#a_g = array.to_device(queue, a, async = False)
print time()-t
def from_array(cls, arr, *args, **kwargs):
ctx = get_device().context
if not arr.ndim in [2, 3, 4]:
raise ValueError(
"dimension of array wrong, should be 1...4 but is %s" %
arr.ndim)
elif arr.ndim == 4:
num_channels = arr.shape[-1]
else:
num_channels = 1
if arr.dtype.type == np.complex64:
num_channels = 2
res = OCLImage.empty(arr.shape,
dtype=np.float32,
num_channels=num_channels)
res.write_array(arr)
res.dtype = np.float32
else:
res = cl.image_from_array(ctx,
prepare(arr),
num_channels=num_channels,
*args,
**kwargs)
res.dtype = arr.dtype
res.num_channels = num_channels
return res
def empty(cls, shape, dtype, num_channels=1, channel_order=None):
ctx = get_device().context
if not len(shape) in [2, 3]:
raise ValueError(
"number of dimension = %s not supported (can be 2 or 3)" %
len(shape))
if not num_channels in [1, 2, 3, 4]:
raise ValueError(
"number of channels = %s not supported (can be 1,2, 3 or 4)" %
num_channels)
mem_flags = cl.mem_flags.READ_WRITE
channel_type = cl.DTYPE_TO_CHANNEL_TYPE[np.dtype(dtype)]
_dict_channel_order = {
1: cl.channel_order.R,
2: cl.channel_order.RG,
3: cl.channel_order.RGB,
4: cl.channel_order.RGBA
}
if channel_order is None:
channel_order = _dict_channel_order[num_channels]
fmt = cl.ImageFormat(channel_order, channel_type)
res = cls(ctx, mem_flags, fmt, shape=shape[::-1])
res.dtype = dtype
res.num_channels = num_channels
return res
def _fft_convolve_numpy(data, h, plan = None,
kernel_is_fft = False,
kernel_is_fftshifted = False):
""" convolving via opencl fft for numpy arrays
data and h must have the same size
"""
dev = get_device()
if data.shape != h.shape:
raise ValueError("data and kernel must have same size! %s vs %s "%(str(data.shape),str(h.shape)))
data_g = OCLArray.from_array(data.astype(np.complex64))
if not kernel_is_fftshifted:
h = np.fft.fftshift(h)
h_g = OCLArray.from_array(h.astype(np.complex64))
res_g = OCLArray.empty_like(data_g)
_fft_convolve_gpu(data_g,h_g,res_g = res_g,
plan = plan,
kernel_is_fft = kernel_is_fft)
res = abs(res_g.get())
del data_g
del h_g
del res_g
return res
def copy_image(self,img, **kwargs):
queue = get_device().queue
return cl.enqueue_copy(queue, self, img,
src_origin = (0,0),
dest_origin = (0,0),
region = self.shape,
**kwargs)
def _ocl_fft_gpu_inplace(ocl_arr, inverse=False, plan=None):
assert_bufs_type(np.complex64, ocl_arr)
if plan is None:
plan = Plan(ocl_arr.shape, queue=get_device().queue)
plan.execute(ocl_arr.data, ocl_arr.data, inverse=inverse)
def _ocl_fft_gpu_inplace(ocl_arr,inverse = False, plan = None):
assert_bufs_type(np.complex64,ocl_arr)
if plan is None:
plan = Plan(ocl_arr.shape, queue = get_device().queue)
plan.execute(ocl_arr.data,ocl_arr.data, inverse = inverse)
def func_to_time(im_size = 256, psf_size = 5, Niter = 10, is_pad = False ):
d = zeros((im_size,)*2,np.float32)
if is_pad:
h = ones((im_size,)*2,np.float32)
else:
h = ones((psf_size,)*2,np.float32)
d_g = gputools.OCLArray.from_array(d)
h_g = gputools.OCLArray.from_array(h)
gputools.get_device().queue.finish()
t = time()
for _ in range(Niter):
f(d_g, h_g)
gputools.get_device().queue.finish()
return (time()-t)/Niter
def _setup_gpu(self):
dev = get_device()
self._queue = dev.queue
self._ctx = dev.context
prog = OCLProgram(absPath("kernels/bpm_3d_kernels.cl"))
# the buffers/ images
Nx, Ny = self.simul_xy
Nx0, Ny0 = self.shape[:2]
self._plan = fft_plan((Ny, Nx), **self.fftplan_kwargs)
self._buf_plane = OCLArray.empty((Ny, Nx), np.complex64)
self._buf_H = OCLArray.empty((Ny, Nx), np.complex64)
self._img_xy = OCLImage.empty((Ny, Nx),
dtype=np.float32,
num_channels=2)
# buffer for the weighted dn average
self.intens_g = OCLArray.empty((1, Ny, Nx), dtype=Bpm3d._real_type)
self.intens_dn_g = OCLArray.empty((1, Ny, Nx), dtype=Bpm3d._real_type)
self.intens_sum_g = OCLArray.zeros((), dtype=Bpm3d._real_type)
self.intens_dn_sum_g = OCLArray.zeros((), dtype=Bpm3d._real_type)
# the kernels
self._kernel_compute_propagator = prog.compute_propagator
self._kernel_compute_propagator.set_scalar_arg_dtypes((None, ) +
(np.float32, ) *
5)
self._kernel_compute_propagator_buf = prog.compute_propagator_buf
self._kernel_compute_propagator_buf.set_scalar_arg_dtypes(
(None, ) + (np.float32, ) * 5 + (None, ) * 2)
self._kernel_mult_complex = prog.mult
self._kernel_im_to_buf_field = prog.img_to_buf_field
self._kernel_im_to_buf_intensity = prog.img_to_buf_intensity
self._kernel_im_to_im_intensity = prog.img_to_img_intensity
self._kernel_buf_to_buf_field = prog.buf_to_buf_field
self._kernel_buf_to_buf_intensity = prog.buf_to_buf_intensity
self._kernel_mult_dn_img_float = prog.mult_dn_image
self._kernel_mult_dn_buf_float = prog.mult_dn
self._kernel_mult_dn_img_complex = prog.mult_dn_image_complex
self._kernel_mult_dn_buf_complex = prog.mult_dn_complex
self._kernel_mult_dn_img_float_local = prog.mult_dn_image_local
self._kernel_mult_dn_buf_float_local = prog.mult_dn_local
self._kernel_mult_dn_img_complex_local = prog.mult_dn_image_complex_local
self._kernel_mult_dn_buf_complex_local = prog.mult_dn_complex_local
self._kernel_reduction = OCLMultiReductionKernel(
np.float32,
neutral="0",
reduce_expr="a+b",
map_exprs=["a[i]", "b[i]"],
arguments="__global float *a, __global float *b")
self._fill_propagator(self.n0)
def copy_image(self, img, **kwargs):
queue = get_device().queue
return cl.enqueue_copy(queue,
self,
img,
src_origin=(0, ) * len(self.imshape()),
dest_origin=(0, ) * len(self.imshape()),
region=self.shape,
**kwargs)
def copy_image(self, img, **kwargs):
queue = get_device().queue
return cl.enqueue_copy(queue,
self.data,
img,
offset=0,
origin=(0, ) * len(img.shape),
region=img.shape,
**kwargs)
def copy_image(self, img, **kwargs):
queue = get_device().queue
return pyopencl.enqueue_copy(queue,
self,
img,
src_origin=(0, 0),
dest_origin=(0, 0),
region=img.shape,
**kwargs)
def copy_buffer(self, buf, **kwargs):
queue = get_device().queue
self.dtype = buf.dtype
return cl.enqueue_copy(queue,
self,
buf.data,
offset=0,
origin=(0, ) * len(self.imshape()),
region=self.imshape(),
**kwargs)
def copy_buffer(self, buf, **kwargs):
queue = get_device().queue
self.dtype = buf.dtype
return pyopencl.enqueue_copy(queue,
self,
buf.data,
offset=0,
origin=(0, 0),
region=self.shape,
**kwargs)
def _integral3_buf(x_g, res_g = None, tmp_g = None):
if not x_g.dtype.type in _output_type_dict:
raise ValueError("dtype %s currently not supported! (%s)" % (x_g.dtype.type, str(_output_type_dict.keys())))
dtype_out = _output_type_dict[x_g.dtype.type]
cl_dtype_in = cl_buffer_datatype_dict[x_g.dtype.type]
cl_dtype_out = cl_buffer_datatype_dict[dtype_out]
dtype_itemsize = np.dtype(dtype_out).itemsize
max_local_size = get_device().get_info("MAX_WORK_GROUP_SIZE")
prog = OCLProgram(abspath("kernels/integral_image.cl"),
build_options=["-D", "DTYPE=%s" % cl_dtype_out])
if x_g.dtype.type != dtype_out:
x_g = x_g.astype(dtype_out)
if tmp_g is None:
tmp_g = OCLArray.empty(x_g.shape, dtype_out)
if res_g is None:
res_g = OCLArray.empty(x_g.shape, dtype_out)
assert_bufs_type(dtype_out, tmp_g, res_g)
nz, ny, nx = x_g.shape
def _scan_single(src, dst, ns, strides):
nx, ny, nz = ns
stride_x, stride_y, stride_z = strides
loc = min(next_power_of_2(nx // 2), max_local_size // 2)
nx_block = 2 * loc
nx_pad = math.ceil(nx / nx_block) * nx_block
nblocks = math.ceil(nx_pad // 2 / loc)
sum_blocks = OCLArray.empty((nz, ny, nblocks), dst.dtype)
shared = cl.LocalMemory(2 * dtype_itemsize * loc)
for b in range(nblocks):
offset = b * loc
prog.run_kernel("scan3d", (loc, ny, nz), (loc, 1, 1),
src.data, dst.data, sum_blocks.data, shared,
np.int32(nx_block),
np.int32(stride_x), np.int32(stride_y), np.int32(stride_z), np.int32(offset), np.int32(b),
np.int32(nblocks), np.int32(ny), np.int32(nx))
if nblocks > 1:
_scan_single(sum_blocks, sum_blocks, (nblocks, ny, nz), (1, nblocks, nblocks * ny))
prog.run_kernel("add_sums3d", (nx_pad, ny, nz), (nx_block, 1, 1),
sum_blocks.data, dst.data,
np.int32(stride_x), np.int32(stride_y), np.int32(stride_z),
np.int32(nblocks), np.int32(ny), np.int32(nx))
_scan_single(x_g, res_g, (nx, ny, nz), (1, nx, nx * ny))
_scan_single(res_g, tmp_g, (ny, nx, nz), (nx, 1, nx * ny))
_scan_single(tmp_g, res_g, (nz, nx, ny), (ny * nx, 1, nx))
return res_g
def test_integral():
max_size = get_device().get_info("GLOBAL_MEM_SIZE") // 32
ndims = (2, 3)
ns = (33, 197, 2183)
dtypes = (np.uint8, np.uint16, np.int32, np.float32)
for dtype, ndim in product(dtypes, ndims):
for shape0 in combinations(ns, ndim):
for shape in permutations(shape0):
if np.prod(shape) > max_size:
continue
single_test(shape, dtype, check=True)
def bench(description,
dshape,
dtype,
func_cpu,
func_gpu,
func_gpu_notransfer=None,
niter=2):
x = np.random.randint(0, 100, dshape).astype(dtype)
func_cpu(x)
t_cpu = time()
for _ in range(niter):
y = func_cpu(x)
t_cpu = (time() - t_cpu) / niter
func_gpu(x)
t_gpu = time()
for _ in range(niter):
y = func_gpu(x)
t_gpu = (time() - t_gpu) / niter
if func_gpu_notransfer is not None:
x_g = OCLArray.from_array(x)
tmp_g = OCLArray.empty_like(x)
func_gpu_notransfer(x_g, tmp_g)
get_device().queue.finish()
t_gpu_notransfer = time()
for _ in range(niter):
func_gpu_notransfer(x_g, tmp_g)
get_device().queue.finish()
t_gpu_notransfer = (time() - t_gpu_notransfer) / niter
else:
t_gpu_notransfer = None
# print("%s\t\t %s\t%d ms \t %d ms"%(description,dshape, 1000*t1,1000*t2))
print("%s| %s %s | %d ms | %d ms | %s" %
(description, dshape, type_name_dict[dtype], 1000 * t_cpu,
1000 * t_gpu, "%d ms" %
(1000 * t_gpu_notransfer) if t_gpu_notransfer is not None else "-"))
return t_cpu, t_gpu, t_gpu_notransfer
def _ocl_fft_numpy(arr,inverse = False, plan = None):
if plan is None:
plan = Plan(arr.shape, queue = get_device().queue)
if arr.dtype != np.complex64:
logger.info("converting %s to complex64, might slow things down..."%arr.dtype)
ocl_arr = OCLArray.from_array(arr.astype(np.complex64,copy=False))
_ocl_fft_gpu_inplace(ocl_arr, inverse = inverse, plan = plan)
return ocl_arr.get()
def time_multi(N, nargs, niter =100):
map_exprs=["%s*x%s[i]"%(i,i) for i in xrange(nargs)]
arguments = ",".join("__global float *x%s"%i for i in xrange(nargs))
k = OCLReductionKernel2(np.float32,
neutral="0", reduce_expr="a+b",
map_exprs=map_exprs,
arguments=arguments)
ins = [OCLArray.from_array(np.ones(N,np.float32)) for _ in xrange(len(map_exprs))]
outs = [OCLArray.empty(1,np.float32) for _ in xrange(len(map_exprs))]
from time import time
t = time()
for _ in xrange(niter):
k(*ins, outs = outs)
get_device().queue.finish()
t = (time()-t)/niter
print "multi reduction: result =", [float(out.get()) for out in outs]
print "multi reduction:\t\t%.2f ms"%(1000*t)
return t
def test_speed_multipass():
import time
from gputools import get_device
N = 256
x = np.linspace(-1, 1, N)
Z, Y, X = np.meshgrid(x, x, x, indexing="ij")
R = np.sqrt(X ** 2 + Y ** 2 + Z ** 2)
d = 200 * np.exp(-10 * R ** 2)
rend = VolumeRenderer((800,) * 2)
rend.set_modelView(mat4_translate(0, 0, -10.))
rend.set_data(d.astype(np.float32))
get_device().queue.finish()
for niter in range(1, 10):
get_device().queue.finish()
t = time.time()
rend.render(method="max_project", maxVal=200.,
currentPart=0, numParts=niter)
get_device().queue.finish()
print("time to render with %s substeps:\t %.2f ms" % (niter, 1000 * (time.time() - t)))
return rend
def test_speed_multipass():
import time
from gputools import get_device
N = 256
x = np.linspace(-1, 1, N)
Z, Y, X = np.meshgrid(x, x, x, indexing="ij")
R = np.sqrt(X**2 + Y**2 + Z**2)
d = 200 * np.exp(-10 * R**2)
rend = VolumeRenderer((800, ) * 2)
rend.set_modelView(mat4_translate(0, 0, -10.))
rend.set_data(d.astype(np.float32))
get_device().queue.finish()
for niter in range(1, 10):
get_device().queue.finish()
t = time.time()
rend.render(method="max_project",
maxVal=200.,
currentPart=0,
numParts=niter)
get_device().queue.finish()
print("time to render with %s substeps:\t %.2f ms" %
(niter, 1000 * (time.time() - t)))
return rend
def from_array(cls,arr, *args, **kwargs):
ctx = get_device().context
if not arr.ndim in [1,2,3,4]:
raise ValueError("dimension of array wrong, should be 1...4 but is %s"%arr.ndim)
elif arr.ndim == 4:
num_channels = arr.shape[-1]
else:
num_channels = None
res = pyopencl.image_from_array(ctx, arr,num_channels = num_channels,
*args, **kwargs)
res.dtype = arr.dtype
return res
def _ocl_fft_numpy(arr, inverse=False, plan=None):
if plan is None:
plan = Plan(arr.shape, queue=get_device().queue)
if arr.dtype != np.complex64:
logger.info("converting %s to complex64, might slow things down..." %
arr.dtype)
ocl_arr = OCLArray.from_array(arr.astype(np.complex64, copy=False))
_ocl_fft_gpu_inplace(ocl_arr, inverse=inverse, plan=plan)
return ocl_arr.get()
def _ocl_fft_gpu(ocl_arr,res_arr = None,inverse = False, plan = None):
assert_bufs_type(np.complex64,ocl_arr)
if plan is None:
plan = Plan(ocl_arr.shape, queue = get_device().queue)
if res_arr is None:
res_arr = OCLArray.empty(ocl_arr.shape,np.complex64)
plan.execute(ocl_arr.data,res_arr.data, inverse = inverse)
return res_arr
def _ocl_fft_gpu(ocl_arr, res_arr=None, inverse=False, plan=None):
assert_bufs_type(np.complex64, ocl_arr)
if plan is None:
plan = Plan(ocl_arr.shape, queue=get_device().queue)
if res_arr is None:
res_arr = OCLArray.empty(ocl_arr.shape, np.complex64)
plan.execute(ocl_arr.data, res_arr.data, inverse=inverse)
return res_arr
def __init__(self,file_name = None,src_str = None, build_options =[], dev = None):
if file_name is not None:
with open(file_name,"r") as f:
src_str = f.read()
if src_str is None:
raise ValueError("empty src_str! ")
if dev is None:
dev = get_device()
self.dev = dev
super(OCLProgram,self).__init__(self.dev.context,src_str)
self.build(options = build_options)
def _setup_gpu(self):
dev = get_device()
self._queue = dev.queue
self._ctx = dev.context
prog = OCLProgram(absPath("kernels/bpm_3d_kernels.cl"))
# the buffers/ images
Nx, Ny = self.simul_xy
Nx0, Ny0 = self.shape[:2]
self._plan = fft_plan((Ny, Nx), **self.fftplan_kwargs)
self._buf_plane = OCLArray.empty((Ny, Nx), np.complex64)
self._buf_H = OCLArray.empty((Ny, Nx), np.complex64)
self._img_xy = OCLImage.empty((Ny, Nx), dtype=np.float32, num_channels=2)
# buffer for the weighted dn average
self.intens_g = OCLArray.empty((1, Ny, Nx), dtype=Bpm3d._real_type)
self.intens_dn_g = OCLArray.empty((1, Ny, Nx), dtype=Bpm3d._real_type)
self.intens_sum_g = OCLArray.zeros((), dtype=Bpm3d._real_type)
self.intens_dn_sum_g = OCLArray.zeros((), dtype=Bpm3d._real_type)
# the kernels
self._kernel_compute_propagator = prog.compute_propagator
self._kernel_compute_propagator.set_scalar_arg_dtypes((None,)+(np.float32,)*5)
self._kernel_compute_propagator_buf = prog.compute_propagator_buf
self._kernel_compute_propagator_buf.set_scalar_arg_dtypes((None,)+(np.float32,)*5+(None,)*2)
self._kernel_mult_complex = prog.mult
self._kernel_im_to_buf_field = prog.img_to_buf_field
self._kernel_im_to_buf_intensity = prog.img_to_buf_intensity
self._kernel_im_to_im_intensity = prog.img_to_img_intensity
self._kernel_buf_to_buf_field = prog.buf_to_buf_field
self._kernel_buf_to_buf_intensity = prog.buf_to_buf_intensity
self._kernel_mult_dn_img_float = prog.mult_dn_image
self._kernel_mult_dn_buf_float = prog.mult_dn
self._kernel_mult_dn_img_complex = prog.mult_dn_image_complex
self._kernel_mult_dn_buf_complex = prog.mult_dn_complex
self._kernel_mult_dn_img_float_local = prog.mult_dn_image_local
self._kernel_mult_dn_buf_float_local = prog.mult_dn_local
self._kernel_mult_dn_img_complex_local = prog.mult_dn_image_complex_local
self._kernel_mult_dn_buf_complex_local = prog.mult_dn_complex_local
self._kernel_reduction = OCLMultiReductionKernel(np.float32,
neutral="0", reduce_expr="a+b",
map_exprs=["a[i]", "b[i]"],
arguments="__global float *a, __global float *b")
self._fill_propagator(self.n0)
def get(self, **kwargs):
queue = get_device().queue
if hasattr(self,"shape"):
imshape = self.shape
else:
imshape = (self.width,)
dshape = imshape[::-1]
if self.format.channel_count>1:
dshape += (self.format.channel_count,)
out = np.empty(dshape,dtype=self.dtype)
pyopencl.enqueue_read_image(queue,self,[0]*len(dshape),imshape,out)
return out.reshape(dshape)
def __init__(self, dtype_out,
neutral, reduce_expr, arguments=None,
map_exprs=[None],
name="reduce_kernel", options=[], preamble=""):
ctx = get_device().context
dtype_out = self.dtype_out = np.dtype(dtype_out)
max_group_size = None
trip_count = 0
self.n_exprs = len(map_exprs)
assert self.n_exprs>0
while True:
self.stage_1_inf = get_reduction_kernel(1, ctx,
dtype_out,
neutral, reduce_expr, arguments,
name=name+"_stage1", options=options, preamble=preamble,
map_exprs=map_exprs,
max_group_size=max_group_size)
kernel_max_wg_size = self.stage_1_inf.kernel.get_work_group_info(
cl.kernel_work_group_info.WORK_GROUP_SIZE,
ctx.devices[0])
if self.stage_1_inf.group_size<=kernel_max_wg_size:
break
else:
max_group_size = kernel_max_wg_size
trip_count += 1
assert trip_count<=2
self.stage_2_inf = get_reduction_kernel(2, ctx,
dtype_out,
neutral, reduce_expr, arguments=arguments,
name=name+"_stage2", options=options,
map_exprs=map_exprs,
preamble=preamble,
max_group_size=max_group_size)
from pytools import any
from pyopencl.tools import VectorArg
assert any(
isinstance(arg_tp, VectorArg)
for arg_tp in self.stage_1_inf.arg_types), \
"ReductionKernel can only be used with functions " \
"that have at least one vector argument"
def get(self, **kwargs):
queue = get_device().queue
imshape = self.imshape()
dshape = imshape[::-1]
ndim = len(imshape)
if self.num_channels > 1:
dshape += (self.num_channels, )
# dshape = (self.num_channels,) + dshape
out = np.empty(dshape, dtype=self.dtype)
#cl.enqueue_read_image(queue, self, [0] * ndim, imshape, out)
cl.enqueue_copy(queue, out, self, origin=(0, ) * ndim, region=imshape)
return out
def fft_plan(shape, dtype=np.complex64, axes=None, fast_math=True):
"""returns an reikna plan/FFT obj of shape dshape
"""
# if not axes is None and any([a<0 for a in axes]):
# raise NotImplementedError("indices of axes have to be non negative, but are: %s"%str(axes))
axes = _convert_axes_to_absolute(shape, axes)
mock_buffer = MockBuffer(dtype, shape)
fft_plan = FFT(mock_buffer, axes=axes).compile(cluda.ocl_api().Thread(get_device().queue),
fast_math=fast_math)
return fft_plan
def __init__(self, dtype_out,
neutral, reduce_expr, arguments=None,
map_exprs=[None],
name="reduce_kernel", options=[], preamble=""):
ctx = get_device().context
dtype_out = self.dtype_out = np.dtype(dtype_out)
max_group_size = None
trip_count = 0
self.n_exprs = len(map_exprs)
assert self.n_exprs>0
while True:
self.stage_1_inf = get_reduction_kernel(1, ctx,
dtype_out,
neutral, reduce_expr, arguments,
name=name+"_stage1", options=options, preamble=preamble,
map_exprs=map_exprs,
max_group_size=max_group_size)
kernel_max_wg_size = self.stage_1_inf.kernel.get_work_group_info(
cl.kernel_work_group_info.WORK_GROUP_SIZE,
ctx.devices[0])
if self.stage_1_inf.group_size<=kernel_max_wg_size:
break
else:
max_group_size = kernel_max_wg_size
trip_count += 1
assert trip_count<=2
self.stage_2_inf = get_reduction_kernel(2, ctx,
dtype_out,
neutral, reduce_expr, arguments=arguments,
name=name+"_stage2", options=options,
map_exprs=map_exprs,
preamble=preamble,
max_group_size=max_group_size)
from pytools import any
from pyopencl.tools import VectorArg
assert any(
isinstance(arg_tp, VectorArg)
for arg_tp in self.stage_1_inf.arg_types), \
"ReductionKernel can only be used with functions " \
"that have at least one vector argument"
def test_time_to_render():
import time
from gputools import get_device
get_device().print_info()
N = 256
x = np.linspace(-1, 1, N)
Z, Y, X = np.meshgrid(x, x, x, indexing="ij")
R = np.sqrt(X**2+Y**2+Z**2)
d = 10000*np.exp(-10*R**2)
rend = VolumeRenderer((600, 600))
rend.set_modelView(mat4_translate(0, 0, -10.))
# rend.set_box_boundaries(.3*np.array([-1,1,-1,1,-1,1]))
t1 = time.time()
get_device().queue.finish()
rend.set_data(d, autoConvert=True)
get_device().queue.finish()
t2 = time.time()
get_device().queue.finish()
rend.render(maxVal=10000.)
out = rend.output
get_device().queue.finish()
print("time to set data %s^3:\t %.2f ms"%(N, 1000*(t2-t1)))
print("time to render %s^3:\t %.2f ms"%(N, 1000*(time.time()-t2)))
return d, rend, out
def test_time_to_render():
import time
from gputools import get_device
get_device().print_info()
N = 256
x = np.linspace(-1, 1, N)
Z, Y, X = np.meshgrid(x, x, x, indexing="ij")
R = np.sqrt(X**2 + Y**2 + Z**2)
d = 10000 * np.exp(-10 * R**2)
rend = VolumeRenderer((600, 600))
rend.set_modelView(mat4_translate(0, 0, -10.))
# rend.set_box_boundaries(.3*np.array([-1,1,-1,1,-1,1]))
t1 = time.time()
get_device().queue.finish()
rend.set_data(d, autoConvert=True)
get_device().queue.finish()
t2 = time.time()
get_device().queue.finish()
rend.render(maxVal=10000.)
out = rend.output
get_device().queue.finish()
print("time to set data %s^3:\t %.2f ms" % (N, 1000 * (t2 - t1)))
print("time to render %s^3:\t %.2f ms" % (N, 1000 * (time.time() - t2)))
return d, rend, out
def get(self, **kwargs):
queue = get_device().queue
if hasattr(self, "shape"):
imshape = self.shape
else:
imshape = (self.width, )
dshape = imshape[::-1]
if self.format.channel_count > 1:
dshape += (self.format.channel_count, )
out = np.empty(dshape, dtype=self.dtype)
pyopencl.enqueue_read_image(queue, self, [0] * len(dshape), imshape,
out)
return out.reshape(dshape)
def time_simple(N, nargs, niter =100):
from gputools import OCLReductionKernel
map_exprs=["%s*x[i]"%i for i in xrange(nargs)]
ks = [OCLReductionKernel(np.float32,
neutral="0", reduce_expr="a+b",
map_expr="%s*x[i]"%i,
arguments="__global float *x") for i in xrange(len(map_exprs))]
ins = [OCLArray.from_array(np.ones(N,np.float32)) for _ in xrange(len(map_exprs))]
outs = [OCLArray.empty(1,np.float32) for _ in xrange(len(map_exprs))]
from time import time
t = time()
for _ in xrange(niter):
for k,inn,out in zip(ks,ins,outs):
k(inn, out = out)
get_device().queue.finish()
t = (time()-t)/niter
print "simple reduction: result =", [float(out.get()) for out in outs]
print "simple reduction:\t\t%.2f ms"%(1000*t)
return t
def copy_buffer(self, buf):
"""
copy content of buf into im
"""
queue = get_device().queue
imshape = self.imshape()
assert imshape == buf.shape[::-1]
ndim = len(imshape)
cl.enqueue_copy(queue,
self,
buf.data,
offset=0,
origin=(0, ) * ndim,
region=imshape)
def _fft_convolve_gpu(data_g, h_g, res_g = None,
plan = None, inplace = False,
kernel_is_fft = False):
""" fft convolve for gpu buffer
"""
_complex_multiply_kernel = OCLElementwiseKernel(
"cfloat_t *a, cfloat_t * b",
"a[i] = cfloat_mul(b[i],a[i])","mult")
dev = get_device()
assert_bufs_type(np.complex64,data_g,h_g)
if data_g.shape != h_g.shape:
raise ValueError("data and kernel must have same size! %s vs %s "%(str(data_g.shape),str(h_g.shape)))
if plan is None:
plan = fft_plan(data_g.shape)
if inplace:
res_g = data_g
else:
if res_g is None:
res_g = OCLArray.empty(data_g.shape,data_g.dtype)
res_g.copy_buffer(data_g)
if not kernel_is_fft:
kern_g = OCLArray.empty(h_g.shape,h_g.dtype)
kern_g.copy_buffer(h_g)
fft(kern_g,inplace=True, plan = plan)
else:
kern_g = h_g
fft(res_g,inplace=True, plan = plan)
#multiply in fourier domain
_complex_multiply_kernel(res_g,kern_g)
fft(res_g,inplace = True, inverse = True, plan = plan)
return res_g
def get(self, **kwargs):
queue = get_device().queue
if hasattr(self,"shape"):
imshape = self.shape
else:
imshape = (self.width,)
dshape = imshape[::-1]
ndim = len(imshape)
if self.num_channels>1:
dshape += (self.num_channels,)
#dshape = (self.num_channels,) + dshape
out = np.empty(dshape,dtype=self.dtype)
cl.enqueue_read_image(queue,self,[0]*ndim,imshape,out)
return out
def copy_buffer(self,buf):
"""
copy content of buf into im
"""
queue = get_device().queue
if hasattr(self,"shape"):
imshape = self.shape
else:
imshape = (self.width,)
assert imshape == buf.shape[::-1]
ndim = len(imshape)
cl.enqueue_copy(queue,self,buf.data,
offset = 0,
origin = (0,)*ndim,
region = imshape)
def write_array(self, data):
queue = get_device().queue
# 1d images dont have a shape but only a width
if hasattr(self,"shape"):
imshape = self.shape
else:
imshape = (self.width,)
ndim = len(imshape)
dshape = data.shape
# if clImg.format.channel_order in [cl.channel_order.RGBA,
# cl.channel_order.BGRA]:
# dshape = dshape[:-1]
if dshape != imshape[::-1]:
raise ValueError("write_array: wrong shape!",data.shape[::-1],imshape)
else:
pyopencl.enqueue_write_image(queue,self,[0]*ndim,imshape,data)
def empty(cls,shape,dtype, num_channels = 1, channel_order = None):
ctx = get_device().context
if not len(shape) in [1,2,3]:
raise ValueError("dimension of shape wrong, should be 1...3 but is %s"%len(shape))
mem_flags = pyopencl.mem_flags.READ_WRITE
channel_type = pyopencl.DTYPE_TO_CHANNEL_TYPE[np.dtype(dtype)]
_dict_channel_order = {1:pyopencl.channel_order.R,
2:pyopencl.channel_order.RG,
3:pyopencl.channel_order.RGB,
4:pyopencl.channel_order.RGBA}
if channel_order is None:
channel_order = _dict_channel_order[num_channels]
fmt = pyopencl.ImageFormat(channel_order, channel_type)
res = pyopencl.Image(ctx, mem_flags,fmt, shape = shape[::-1])
res.dtype = dtype
return res
def _convolve3_old(data,h, dev = None):
"""convolves 3d data with kernel h on the GPU Device dev
boundary conditions are clamping to edge.
h is converted to float32
if dev == None the default one is used
"""
if dev is None:
dev = get_device()
if dev is None:
raise ValueError("no OpenCLDevice found...")
dtype = data.dtype.type
dtypes_options = {np.float32:"",
np.uint16:"-D SHORTTYPE"}
if not dtype in dtypes_options.keys():
raise TypeError("data type %s not supported yet, please convert to:"%dtype,dtypes_options.keys())
prog = OCLProgram(abspath("kernels/convolve3.cl"),
build_options = dtypes_options[dtype])
hbuf = OCLArray.from_array(h.astype(np.float32))
img = OCLImage.from_array(data)
res = OCLArray.empty(data.shape,dtype=np.float32)
Ns = [np.int32(n) for n in data.shape+h.shape]
prog.run_kernel("convolve3d",img.shape,None,
img,hbuf.data,res.data,
*Ns)
return res.get()
def from_array(cls,arr, *args, **kwargs):
ctx = get_device().context
if not arr.ndim in [2,3,4]:
raise ValueError("dimension of array wrong, should be 1...4 but is %s"%arr.ndim)
elif arr.ndim == 4:
num_channels = arr.shape[-1]
else:
num_channels = None
if arr.dtype.type == np.complex64:
num_channels = 2
res = OCLImage.empty(arr.shape,dtype = np.float32, num_channels=num_channels)
res.write_array(arr)
res.dtype = np.float32
else:
res = cl.image_from_array(ctx, arr,num_channels = num_channels,
*args, **kwargs)
res.dtype = arr.dtype
res.num_channels = num_channels
return res
def copy_image(self,img, **kwargs):
queue = get_device().queue
return cl.enqueue_copy(queue, self.data, img, offset = 0,
origin = (0,0),region = img.shape,
**kwargs)
def write_array(self,data, **kwargs):
queue = get_device().queue
return cl.enqueue_write_buffer(queue, self.data, data,
**kwargs)
def copy_buffer(self,buf, **kwargs):
queue = get_device().queue
return cl.enqueue_copy(queue, self.data, buf.data,
**kwargs)
