def get_astype_function(dtype_dest, dtype_src, pitch=True):
type_dest = dtype_to_ctype(dtype_dest)
type_src = dtype_to_ctype(dtype_src)
name = "astype"
operation = ""
if pitch:
func = func_compile(
name, pitch_template % {
"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare(
[np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(
name, non_pitch_template % {
"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_conj_function(dtype, pitch=True):
type_src = dtype_to_ctype(dtype)
if dtype == np.complex128:
operation = "pycuda::conj"
elif dtype == np.complex64:
operation = "pycuda::conj"
else:
raise TypeError(
"Only complex arrays are allowed to perform conjugation")
name = "conj"
if pitch:
func = func_compile(
name, pitch_template % {
"name": name,
"dest_type": type_src,
"src_type": type_src,
"operation": operation,
})
func.prepare(
[np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(
name, non_pitch_template % {
"name": name,
"dest_type": type_src,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_divarray_function(left_dtype, right_dtype, rslt_dtype, pitch = True):
type_left = dtype_to_ctype(left_dtype)
type_right = dtype_to_ctype(right_dtype)
type_rslt = dtype_to_ctype(rslt_dtype)
name = "divarray"
operation = "/"
if pitch:
func = func_compile(name, pitch_array_op_template % {"name": name,
"dest_type": type_rslt,
"left_type": type_left,
"right_type": type_right,
"operation": operation,
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(name, non_pitch_array_op_template % {"name": name,
"dest_type": type_rslt,
"left_type": type_left,
"right_type": type_right,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.intp, np.int32])
return func
def get_divarray_function(left_dtype, right_dtype, rslt_dtype, pitch=True):
type_left = dtype_to_ctype(left_dtype)
type_right = dtype_to_ctype(right_dtype)
type_rslt = dtype_to_ctype(rslt_dtype)
name = "divarray"
operation = "/"
if pitch:
func = func_compile(
name, pitch_array_op_template % {
"name": name,
"dest_type": type_rslt,
"left_type": type_left,
"right_type": type_right,
"operation": operation,
})
func.prepare([
np.int32, np.int32, np.intp, np.int32, np.intp, np.int32, np.intp,
np.int32
])
else:
func = func_compile(
name, non_pitch_array_op_template % {
"name": name,
"dest_type": type_rslt,
"left_type": type_left,
"right_type": type_right,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.intp, np.int32])
return func
def get_scalardiv_function(src_type, pitch=True):
type_src = dtype_to_ctype(src_type)
name = "scalardiv"
operation = "/"
if pitch:
func = func_compile(
name, pitch_right_scalar_op_template % {
"name": name,
"dest_type": type_src,
"operation": operation,
})
func.prepare([
np.int32, np.int32, np.intp, np.int32, np.intp, np.int32,
src_type.type
])
else:
func = func_compile(
name, non_pitch_right_scalar_op_template % {
"name": name,
"dest_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, src_type.type, np.int32])
return func
def get_conj_function(dtype, pitch = True):
type_src = dtype_to_ctype(dtype)
if dtype == np.complex128:
operation = "pycuda::conj"
elif dtype == np.complex64:
operation = "pycuda::conj"
else:
raise TypeError("Only complex arrays are allowed to perform conjugation")
name = "conj"
if pitch:
func = func_compile(name, pitch_template % {"name": name,
"dest_type": type_src,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(name, non_pitch_template % {"name": name,
"dest_type": type_src,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_fill_function(dtype, pitch = True):
type_dst = dtype_to_ctype(dtype)
name = "fill"
if pitch:
func = func_compile(name, fill_pitch_template % {"name": name,
"type_dst": type_dst})
func.prepare([np.int32, np.int32, np.intp, np.int32, dtype.type])
else:
func = func_compile(name, fill_nonpitch_template % {"name": name,
"type_dst": type_dst})
func.prepare([np.int32, np.intp, dtype.type])
return func
def get_realimag_function(dtype, real=True, pitch=True):
type_src = dtype_to_ctype(dtype)
if dtype == np.complex64:
type_dest = "float"
if real:
operation = "pycuda::real"
name = "real"
else:
operation = "pycuda::imag"
name = "imag"
elif dtype == np.complex128:
type_dest = "double"
if real:
operation = "pycuda::real"
name = "real"
else:
operation = "pycuda::imag"
name = "imag"
else:
raise TypeError(
"only support complex inputs as numpy.complex64 or numpy.complex128"
)
if pitch:
func = func_compile(
name, pitch_template % {
"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare(
[np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(
name, non_pitch_template % {
"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_scalardiv_function(src_type, pitch = True):
type_src = dtype_to_ctype(src_type)
name = "scalardiv"
operation = "/"
if pitch:
func = func_compile(name, pitch_right_scalar_op_template % {"name": name,
"dest_type": type_src,
"operation": operation,
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32, src_type.type])
else:
func = func_compile(name, non_pitch_right_scalar_op_template % {"name": name,
"dest_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, src_type.type, np.int32])
return func
def get_fill_function(dtype, pitch=True):
type_dst = dtype_to_ctype(dtype)
name = "fill"
if pitch:
func = func_compile(
name, fill_pitch_template % {
"name": name,
"type_dst": type_dst
})
func.prepare([np.int32, np.int32, np.intp, np.int32, dtype.type])
else:
func = func_compile(
name, fill_nonpitch_template % {
"name": name,
"type_dst": type_dst
})
func.prepare([np.int32, np.intp, dtype.type])
return func
def get_abs_function(dtype, pitch=True):
type_src = dtype_to_ctype(dtype)
if dtype == np.complex128:
operation = "pycuda::abs"
type_dest = "double"
elif dtype == np.complex64:
operation = "pycuda::abs"
type_dest = "float"
elif dtype == np.float64:
operation = "fabs"
type_dest = "double"
elif dtype == np.float32:
operation = "fabsf"
type_dest = "float"
else:
operation = "abs"
type_dest = dtype_to_ctype(dtype)
name = "abs_function"
if pitch:
func = func_compile(
name, pitch_template % {
"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare(
[np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(
name, non_pitch_template % {
"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_resize_function(dtype):
type_src = dtype_to_ctype(dtype)
name = "resize"
func = func_compile(name, reshape_template % {"name": name,
"dest_type": type_src,
"src_type": type_src,
"operation": "",
})
func.prepare([np.int32, np.int32, np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
return func
def get_astype_function(dtype_dest, dtype_src, pitch = True):
type_dest = dtype_to_ctype(dtype_dest)
type_src = dtype_to_ctype(dtype_src)
name = "astype"
operation = ""
if pitch:
func = func_compile(name, pitch_template % {"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(name, non_pitch_template % {"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_abs_function(dtype, pitch = True):
type_src = dtype_to_ctype(dtype)
if dtype == np.complex128:
operation = "pycuda::abs"
type_dest = "double"
elif dtype == np.complex64:
operation = "pycuda::abs"
type_dest = "float"
elif dtype == np.float64:
operation = "fabs"
type_dest = "double"
elif dtype == np.float32:
operation = "fabsf"
type_dest = "float"
else:
operation = "abs"
type_dest = dtype_to_ctype(dtype)
name = "abs_function"
if pitch:
func = func_compile(name, pitch_template % {"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(name, non_pitch_template % {"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_transpose_function(dtype, conj = False):
src_type = dtype_to_ctype(dtype)
name = "trans"
operation = ""
if conj:
if dtype == np.complex128:
operation = "pycuda::conj"
elif dtype == np.complex64:
operation = "pycuda::conj"
func = func_compile(name, transpose_template % {"name": name, "type": src_type, "operation": operation})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
return func
def get_realimag_function(dtype, real = True, pitch = True):
type_src = dtype_to_ctype(dtype)
if dtype == np.complex64:
type_dest = "float"
if real:
operation = "pycuda::real"
name = "real"
else:
operation = "pycuda::imag"
name = "imag"
elif dtype == np.complex128:
type_dest = "double"
if real:
operation = "pycuda::real"
name = "real"
else:
operation = "pycuda::imag"
name = "imag"
else:
raise TypeError("only support complex inputs as numpy.complex64 or numpy.complex128")
if pitch:
func = func_compile(name, pitch_template % {"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
else:
func = func_compile(name, non_pitch_template % {"name": name,
"dest_type": type_dest,
"src_type": type_src,
"operation": operation,
})
func.prepare([np.intp, np.intp, np.int32])
return func
def get_resize_function(dtype):
type_src = dtype_to_ctype(dtype)
name = "resize"
func = func_compile(
name, reshape_template % {
"name": name,
"dest_type": type_src,
"src_type": type_src,
"operation": "",
})
func.prepare([
np.int32, np.int32, np.int32, np.int32, np.intp, np.int32, np.intp,
np.int32
])
return func
def get_transpose_function(dtype, conj=False):
src_type = dtype_to_ctype(dtype)
name = "trans"
operation = ""
if conj:
if dtype == np.complex128:
operation = "pycuda::conj"
elif dtype == np.complex64:
operation = "pycuda::conj"
func = func_compile(
name, transpose_template % {
"name": name,
"type": src_type,
"operation": operation
})
func.prepare([np.int32, np.int32, np.intp, np.int32, np.intp, np.int32])
return func