def _compile_out_x(self):
# 2D: int out_x_0 = i / (out_1) % out_0;
# int out_x_1 = i % out_1;
def aux(out_x, outs):
head = outs[0]
tail = outs[1:]
if tail:
return 'int {} = i / ({}) % {};'.format(
out_x, conv_nd_kernel.mulexp(tail), head)
else:
return 'int {} = i % {};'.format(out_x, head)
out_xs = conv_nd_kernel.vars('out_x', self.ndim)
out_xs_decls = conv_nd_kernel.map_(
aux, out_xs, conv_nd_kernel.succ_sublists(self.outs))
return out_xs_decls, out_xs
def _compile_x(self):
# 2D: int x_0 = i / (d_1) % d_0 + p_0;
# int x_1 = i % d_1 + p_1;
def aux(x, ds, p):
head = ds[0]
tail = ds[1:]
if tail:
return 'int {} = i / ({}) % {} + {};'.format(
x, conv_nd_kernel.mulexp(tail), head, p)
else:
return 'int {} = i % {} + {};'.format(x, head, p)
xs = conv_nd_kernel.vars('x', self.ndim)
xs_decls = conv_nd_kernel.map_(
aux, xs, conv_nd_kernel.succ_sublists(self.ds), self.ps)
return xs_decls, xs
def _compile_out_x(self):
# 2D: int out_x_0 = i / (out_1) % out_0;
# int out_x_1 = i % out_1;
def aux(out_x, outs):
head = outs[0]
tail = outs[1:]
if tail:
return 'int {} = i / ({}) % {};'.format(
out_x, conv_nd_kernel.mulexp(tail), head)
else:
return 'int {} = i % {};'.format(out_x, head)
out_xs = conv_nd_kernel.vars('out_x', self.ndim)
out_xs_decls = conv_nd_kernel.map_(
aux, out_xs, conv_nd_kernel.succ_sublists(self.outs))
return out_xs_decls, out_xs
def _compile_x(self):
# 2D: int x_0 = i / (d_1) % d_0 + p_0;
# int x_1 = i % d_1 + p_1;
def aux(x, ds, p):
head = ds[0]
tail = ds[1:]
if tail:
return 'int {} = i / ({}) % {} + {};'.format(
x, conv_nd_kernel.mulexp(tail), head, p)
else:
return 'int {} = i % {} + {};'.format(x, head, p)
xs = conv_nd_kernel.vars('x', self.ndim)
xs_decls = conv_nd_kernel.map_(
aux, xs, conv_nd_kernel.succ_sublists(self.ds), self.ps)
return xs_decls, xs
def _compile_max_x(self):
def aux(max_val, out_val, stride_val, pad_val, ksize_vals):
head = ksize_vals[0]
tail = ksize_vals[1:]
if tail:
command = 'int {} = max(0, {} * {} - {} + index / ({}) % {});'
return command.format(max_val, out_val, stride_val, pad_val,
conv_nd_kernel.mulexp(tail), head)
else:
return 'int {} = max(0, {} * {} - {} + index % {});'.format(
max_val, out_val, stride_val, pad_val, head)
max_vals = conv_nd_kernel.vars('max', self.ndim)
out_vals = conv_nd_kernel.vars('out_x', self.ndim)
stride_vals = conv_nd_kernel.vars('s', self.ndim)
pad_vals = conv_nd_kernel.vars('p', self.ndim)
ksize_vals = conv_nd_kernel.vars('k', self.ndim)
offset_ks_decls = conv_nd_kernel.map_(
aux, max_vals, out_vals, stride_vals, pad_vals,
conv_nd_kernel.succ_sublists(ksize_vals))
return offset_ks_decls
def _compile_max_x(self):
def aux(max_val, out_val, stride_val, pad_val, ksize_vals):
head = ksize_vals[0]
tail = ksize_vals[1:]
if tail:
command = 'int {} = max(0, {} * {} - {} + index / ({}) % {});'
return command.format(
max_val, out_val, stride_val, pad_val,
conv_nd_kernel.mulexp(tail), head)
else:
return 'int {} = max(0, {} * {} - {} + index % {});'.format(
max_val, out_val, stride_val, pad_val, head)
max_vals = conv_nd_kernel.vars('max', self.ndim)
out_vals = conv_nd_kernel.vars('out_x', self.ndim)
stride_vals = conv_nd_kernel.vars('s', self.ndim)
pad_vals = conv_nd_kernel.vars('p', self.ndim)
ksize_vals = conv_nd_kernel.vars('k', self.ndim)
offset_ks_decls = conv_nd_kernel.map_(
aux, max_vals, out_vals, stride_vals, pad_vals,
conv_nd_kernel.succ_sublists(ksize_vals))
return offset_ks_decls