def __init__(self, dtype): from pycuda.curandom import md5_code from pycuda.elementwise import get_elwise_kernel if dtype == numpy.complex64: self._func = get_elwise_kernel( "float2 *dest, unsigned int seed", md5_code + """ #define POW_2_M32 (1/4294967296.0f) dest[i] = make_float2(a*POW_2_M32, b*POW_2_M32); if ((i += total_threads) < n) dest[i] = make_float2(c*POW_2_M32, d*POW_2_M32); """, "md5_rng_float") elif dtype == numpy.complex128: self._func = get_elwise_kernel( "pycuda::complex<double> *dest, unsigned int seed", md5_code + """ #define POW_2_M32 (1/4294967296.0) #define POW_2_M64 (1/18446744073709551616.) dest[i] = pycuda::complex<double>( a*POW_2_M32 + b*POW_2_M64, c*POW_2_M32 + d*POW_2_M64); """, "md5_rng_float")
def rand(shape, dtype=numpy.float32, stream=None):
from pycuda.gpuarray import GPUArray
from pycuda.elementwise import get_elwise_kernel
result = GPUArray(shape, dtype)
if dtype == numpy.float32:
func = get_elwise_kernel(
"float *dest, unsigned int seed",
md5_code + """
#define POW_2_M32 (1/4294967296.0f)
dest[i] = a*POW_2_M32;
if ((i += total_threads) < n)
dest[i] = b*POW_2_M32;
if ((i += total_threads) < n)
dest[i] = c*POW_2_M32;
if ((i += total_threads) < n)
dest[i] = d*POW_2_M32;
""",
"md5_rng_float")
elif dtype == numpy.float64:
func = get_elwise_kernel(
"double *dest, unsigned int seed",
md5_code + """
#define POW_2_M32 (1/4294967296.0)
#define POW_2_M64 (1/18446744073709551616.)
dest[i] = a*POW_2_M32 + b*POW_2_M64;
if ((i += total_threads) < n)
{
dest[i] = c*POW_2_M32 + d*POW_2_M64;
}
""",
"md5_rng_float")
elif dtype in [numpy.int32, numpy.uint32]:
func = get_elwise_kernel(
"unsigned int *dest, unsigned int seed",
md5_code + """
dest[i] = a;
if ((i += total_threads) < n)
dest[i] = b;
if ((i += total_threads) < n)
dest[i] = c;
if ((i += total_threads) < n)
dest[i] = d;
""",
"md5_rng_int")
else:
raise NotImplementedError;
func.set_block_shape(*result._block)
func.prepared_async_call(result._grid, stream,
result.gpudata, numpy.random.randint(2**31-1), result.size)
return result
def guarded_div_kernel(self, dtype_x, dtype_y, dtype_z):
from pycuda.elementwise import get_elwise_kernel
from pycuda.tools import dtype_to_ctype
return get_elwise_kernel(
"%(tp_x)s *x, %(tp_y)s *y, %(tp_z)s *z" % {
"tp_x": dtype_to_ctype(dtype_x),
"tp_y": dtype_to_ctype(dtype_y),
"tp_z": dtype_to_ctype(dtype_z),
}, "z[i] = y[i] == 0 ? 0 : (x[i] / y[i])", "divide")
def guarded_div_kernel(self, dtype_x, dtype_y, dtype_z):
from pycuda.elementwise import get_elwise_kernel
from pycuda.tools import dtype_to_ctype
return get_elwise_kernel(
"%(tp_x)s *x, %(tp_y)s *y, %(tp_z)s *z" % {
"tp_x": dtype_to_ctype(dtype_x),
"tp_y": dtype_to_ctype(dtype_y),
"tp_z": dtype_to_ctype(dtype_z),
},
"z[i] = y[i] == 0 ? 0 : (x[i] / y[i])",
"divide")
def make_kernel_internal(self, args, instructions):
from pycuda.elementwise import get_elwise_kernel
return get_elwise_kernel(args, instructions, name="vector_expression")
else:
raise ValueError("Incompatible dtype")
return df
def linear(x):
pass
def df_linear(x):
return x
sample_dropout_mask_kernel = get_elwise_kernel(
"float *mat, float *dropout, float dropout_probability",
"""
if (dropout[i] <= dropout_probability) {
dropout[i] = 0.;
mat[i] = 0.;
} else {
dropout[i] = 1.;
}
""",
"sample_dropout_mask")
def sample_dropout_mask(x, dropout_probability=.5, columns=None, stream=None):
""" Samples a dropout mask and applies it in place"""
assert x.flags.c_contiguous
if columns is not None:
assert len(columns) == 2
x_tmp = x
def make_kernel_internal(self, args, instructions):
from pycuda.elementwise import get_elwise_kernel
return get_elwise_kernel(args, instructions, name="vector_expression")