def _convolve_buf(data_g, h_g, res_g=None):
"""
buffer variant
"""
assert_bufs_type(np.float32, data_g, h_g)
prog = OCLProgram(abspath("kernels/convolve.cl"))
if res_g is None:
res_g = OCLArray.empty(data_g.shape, dtype=np.float32)
Nhs = [np.int32(n) for n in h_g.shape]
kernel_name = "convolve%sd_buf" % (len(data_g.shape))
try:
prog.run_kernel(kernel_name, data_g.shape[::-1], None,
data_g.data, h_g.data, res_g.data,
*Nhs)
except cl.cffi_cl.LogicError as e:
# this catches the logicerror if the kernel is to big for constant memory
if e.code == -52:
kernel_name = "convolve%sd_buf_global" % (len(data_g.shape))
prog.run_kernel(kernel_name, data_g.shape[::-1], None,
data_g.data, h_g.data, res_g.data,
*Nhs)
else:
raise e
return res_g
def _filt(data_g, size=(3, 3, 3), res_g=None):
assert_bufs_type(np.float32, data_g)
with open(abspath("kernels/generic_reduce_filter.cl"), "r") as f:
tpl = Template(f.read())
rendered = tpl.render(FSIZE_X=size[-1],
FSIZE_Y=size[-2],
FSIZE_Z=size[-3],
FUNC=FUNC,
DEFAULT=DEFAULT)
prog = OCLProgram(src_str=rendered)
tmp_g = OCLArray.empty_like(data_g)
if res_g is None:
res_g = OCLArray.empty_like(data_g)
prog.run_kernel("filter_3_x", data_g.shape[::-1], None, data_g.data,
res_g.data)
prog.run_kernel("filter_3_y", data_g.shape[::-1], None, res_g.data,
tmp_g.data)
prog.run_kernel("filter_3_z", data_g.shape[::-1], None, tmp_g.data,
res_g.data)
return res_g
def _ocl_fft_gpu(plan, ocl_arr, res_arr=None, inverse=False):
assert_bufs_type(np.complex64, ocl_arr)
if res_arr is None:
res_arr = OCLArray.empty_like(ocl_arr)
plan(ocl_arr, res_arr, inverse=inverse)
return res_arr
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(plan, ocl_arr,res_arr = None, inverse = False, batch = 1):
assert_bufs_type(np.complex64,ocl_arr)
if res_arr is None:
res_arr = OCLArray.empty(ocl_arr.shape,np.complex64)
plan.execute(ocl_arr.data,res_arr.data, inverse = inverse, batch = batch)
return res_arr
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 _max_filter_gpu(data_g, size=5, res_g=None):
assert_bufs_type(np.float32, data_g)
assert (len(data_g.shape) == len(size))
if len(data_g.shape) == 2:
return _filter_max_2_gpu(data_g, size=size, res_g=res_g)
elif len(data_g.shape) == 3:
return _filter_max_3_gpu(data_g, size=size, res_g=res_g)
else:
raise NotImplementedError("only 2 or 3d arrays are supported for now")
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 _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 _filter_max_2_gpu(data_g, size=10, res_g=None):
assert_bufs_type(np.float32, data_g)
prog = OCLProgram(abspath("kernels/minmax_filter.cl"))
tmp_g = OCLArray.empty_like(data_g)
if res_g is None:
res_g = OCLArray.empty_like(data_g)
prog.run_kernel("max_2_x", data_g.shape[::-1], None, data_g.data,
tmp_g.data, np.int32(size[-1]))
prog.run_kernel("max_2_y", data_g.shape[::-1], None, tmp_g.data,
res_g.data, np.int32(size[-2]))
return res_g
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 _convolve_sep2_gpu(data_g, hx_g, hy_g, res_g = None):
assert_bufs_type(np.float32,data_g,hx_g,hy_g)
prog = OCLProgram(abspath("kernels/convolve_sep.cl"))
Ny,Nx = hy_g.shape[0],hx_g.shape[0]
tmp_g = OCLArray.empty_like(data_g)
if res_g is None:
res_g = OCLArray.empty_like(data_g)
prog.run_kernel("conv_sep2_x",data_g.shape[::-1],None,data_g.data,hx_g.data,tmp_g.data,np.int32(Nx))
prog.run_kernel("conv_sep2_y",data_g.shape[::-1],None,tmp_g.data,hy_g.data,res_g.data,np.int32(Ny))
return res_g
def _convolve_sep2_gpu(data_g, hx_g, hy_g, res_g=None):
assert_bufs_type(np.float32, data_g, hx_g, hy_g)
prog = OCLProgram(abspath("kernels/convolve_sep.cl"))
Ny, Nx = hy_g.shape[0], hx_g.shape[0]
tmp_g = OCLArray.empty_like(data_g)
if res_g is None:
res_g = OCLArray.empty_like(data_g)
prog.run_kernel("conv_sep2_x", data_g.shape[::-1], None, data_g.data,
hx_g.data, tmp_g.data, np.int32(Nx))
prog.run_kernel("conv_sep2_y", data_g.shape[::-1], None, tmp_g.data,
hy_g.data, res_g.data, np.int32(Ny))
return res_g
def _fft_convolve_gpu(data_g, h_g, res_g = None,
plan = None, inplace = False,
kernel_is_fft = False):
""" fft convolve for gpu buffer
"""
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
print res_g.dtype, res_g.nbytes
_complex_multiply_kernel(res_g,kern_g)
fft(res_g,inplace = True, inverse = True, plan = plan)
return res_g
def _convolve_buf(data_g, h_g , res_g = None):
"""
buffer variant
"""
assert_bufs_type(np.float32,data_g,h_g)
prog = OCLProgram(abspath("kernels/convolve.cl"))
if res_g is None:
res_g = OCLArray.empty(data_g.shape,dtype=np.float32)
Nhs = [np.int32(n) for n in h_g.shape]
kernel_name = "convolve%sd_buf"%(len(data_g.shape))
prog.run_kernel(kernel_name,data_g.shape[::-1],None,
data_g.data,h_g.data,res_g.data,
*Nhs)
return res_g
def _ocl_fft_gpu_inplace(plan, ocl_arr, inverse=False):
assert_bufs_type(np.complex64, ocl_arr)
plan(ocl_arr, ocl_arr, inverse=inverse)
def _ocl_fft_gpu_inplace(plan, ocl_arr, inverse = False, batch = 1):
assert_bufs_type(np.complex64,ocl_arr)
plan.execute(ocl_arr.data,ocl_arr.data, inverse = inverse, batch = batch)
