def get_kernel_call(self, codegen_state, name, gsize, lsize, extra_args):
ecm = self.get_expression_to_code_mapper(codegen_state)
from pymbolic.mapper.stringifier import PREC_COMPARISON, PREC_NONE
result = []
from cgen import Statement as S, Block
if lsize:
result.append(
S("assert(programCount == %s)"
% ecm(lsize[0], PREC_COMPARISON)))
if gsize:
launch_spec = "[%s]" % ", ".join(
ecm(gs_i, PREC_NONE)
for gs_i in gsize)
else:
launch_spec = ""
arg_names, arg_decls = self._arg_names_and_decls(codegen_state)
result.append(S(
"launch%s %s(%s)" % (
launch_spec,
name,
", ".join(arg_names)
)))
return Block(result)
def get_kernel_call(self, codegen_state, name, gsize, lsize, extra_args):
ecm = self.get_expression_to_code_mapper(codegen_state)
from pymbolic.mapper.stringifier import PREC_NONE
result = []
from cgen import Statement as S, Block
if lsize:
result.append(
S("assert(programCount == (%s))" % ecm(lsize[0], PREC_NONE)))
arg_names, arg_decls = self._arg_names_and_decls(codegen_state)
from cgen.ispc import ISPCLaunch
result.append(
ISPCLaunch(tuple(ecm(gs_i, PREC_NONE) for gs_i in gsize),
"%s(%s)" % (name, ", ".join(arg_names))))
return Block(result)
def generate_assignment_instruction_code(codegen_state, insn):
kernel = codegen_state.kernel
ecm = codegen_state.expression_to_code_mapper
from loopy.expression import VectorizabilityChecker
# {{{ vectorization handling
if codegen_state.vectorization_info:
if insn.atomicity:
raise Unvectorizable("atomic operation")
vinfo = codegen_state.vectorization_info
vcheck = VectorizabilityChecker(kernel, vinfo.iname, vinfo.length)
lhs_is_vector = vcheck(insn.assignee)
rhs_is_vector = vcheck(insn.expression)
if not lhs_is_vector and rhs_is_vector:
raise Unvectorizable("LHS is scalar, RHS is vector, cannot assign")
is_vector = lhs_is_vector
del lhs_is_vector
del rhs_is_vector
# }}}
from pymbolic.primitives import Variable, Subscript, Lookup
from loopy.symbolic import LinearSubscript
lhs = insn.assignee
if isinstance(lhs, Lookup):
lhs = lhs.aggregate
if isinstance(lhs, Variable):
assignee_var_name = lhs.name
assignee_indices = ()
elif isinstance(lhs, Subscript):
assignee_var_name = lhs.aggregate.name
assignee_indices = lhs.index_tuple
elif isinstance(lhs, LinearSubscript):
assignee_var_name = lhs.aggregate.name
assignee_indices = (lhs.index, )
else:
raise RuntimeError("invalid lvalue '%s'" % lhs)
del lhs
result = codegen_state.ast_builder.emit_assignment(codegen_state, insn)
# {{{ tracing
lhs_dtype = codegen_state.kernel.get_var_descriptor(
assignee_var_name).dtype
if kernel.options.trace_assignments or kernel.options.trace_assignment_values:
if codegen_state.vectorization_info and is_vector:
raise Unvectorizable("tracing does not support vectorization")
from pymbolic.mapper.stringifier import PREC_NONE
lhs_code = codegen_state.expression_to_code_mapper(
insn.assignee, PREC_NONE)
from cgen import Statement as S # noqa
gs, ls = kernel.get_grid_size_upper_bounds()
printf_format = "{}.{}[{}][{}]: {}".format(
kernel.name, insn.id, ", ".join("gid%d=%%d" % i
for i in range(len(gs))),
", ".join("lid%d=%%d" % i for i in range(len(ls))),
assignee_var_name)
printf_args = (["gid(%d)" % i for i in range(len(gs))] +
["lid(%d)" % i for i in range(len(ls))])
if assignee_indices:
printf_format += "[%s]" % ",".join(len(assignee_indices) * ["%d"])
printf_args.extend(
ecm(i, prec=PREC_NONE, type_context="i")
for i in assignee_indices)
if kernel.options.trace_assignment_values:
if lhs_dtype.numpy_dtype.kind == "i":
printf_format += " = %d"
printf_args.append(lhs_code)
elif lhs_dtype.numpy_dtype.kind == "f":
printf_format += " = %g"
printf_args.append(lhs_code)
elif lhs_dtype.numpy_dtype.kind == "c":
printf_format += " = %g + %gj"
printf_args.extend(["(%s).x" % lhs_code, "(%s).y" % lhs_code])
if printf_args:
printf_args_str = ", " + ", ".join(str(v) for v in printf_args)
else:
printf_args_str = ""
printf_insn = S('printf("{}\\n"{})'.format(printf_format,
printf_args_str))
from cgen import Block
if kernel.options.trace_assignment_values:
result = Block([result, printf_insn])
else:
# print first, execute later -> helps find segfaults
result = Block([printf_insn, result])
# }}}
return result
def generate_assignment_instruction_code(codegen_state, insn):
kernel = codegen_state.kernel
ecm = codegen_state.expression_to_code_mapper
from loopy.expression import dtype_to_type_context, VectorizabilityChecker
# {{{ vectorization handling
if codegen_state.vectorization_info:
if insn.atomicity:
raise Unvectorizable("atomic operation")
vinfo = codegen_state.vectorization_info
vcheck = VectorizabilityChecker(kernel, vinfo.iname, vinfo.length)
lhs_is_vector = vcheck(insn.assignee)
rhs_is_vector = vcheck(insn.expression)
if not lhs_is_vector and rhs_is_vector:
raise Unvectorizable("LHS is scalar, RHS is vector, cannot assign")
is_vector = lhs_is_vector
del lhs_is_vector
del rhs_is_vector
# }}}
from pymbolic.primitives import Variable, Subscript
from loopy.symbolic import LinearSubscript
lhs = insn.assignee
if isinstance(lhs, Variable):
assignee_var_name = lhs.name
assignee_indices = ()
elif isinstance(lhs, Subscript):
assignee_var_name = lhs.aggregate.name
assignee_indices = lhs.index_tuple
elif isinstance(lhs, LinearSubscript):
assignee_var_name = lhs.aggregate.name
assignee_indices = (lhs.index, )
else:
raise RuntimeError("invalid lvalue '%s'" % lhs)
lhs_var = kernel.get_var_descriptor(assignee_var_name)
lhs_dtype = lhs_var.dtype
if insn.atomicity is not None:
lhs_atomicity = [
a for a in insn.atomicity if a.var_name == assignee_var_name
]
assert len(lhs_atomicity) <= 1
if lhs_atomicity:
lhs_atomicity, = lhs_atomicity
else:
lhs_atomicity = None
else:
lhs_atomicity = None
from loopy.kernel.data import AtomicInit, AtomicUpdate
lhs_code = ecm(insn.assignee, prec=PREC_NONE, type_context=None)
rhs_type_context = dtype_to_type_context(kernel.target, lhs_dtype)
if lhs_atomicity is None:
result = codegen_state.ast_builder.emit_assignment(
codegen_state, lhs_code,
ecm(insn.expression,
prec=PREC_NONE,
type_context=rhs_type_context,
needed_dtype=lhs_dtype))
elif isinstance(lhs_atomicity, AtomicInit):
raise NotImplementedError("atomic init")
elif isinstance(lhs_atomicity, AtomicUpdate):
codegen_state.seen_atomic_dtypes.add(lhs_dtype)
result = codegen_state.ast_builder.generate_atomic_update(
kernel, codegen_state, lhs_atomicity, lhs_var, insn.assignee,
insn.expression, lhs_dtype, rhs_type_context)
else:
raise ValueError("unexpected lhs atomicity type: %s" %
type(lhs_atomicity).__name__)
# {{{ tracing
if kernel.options.trace_assignments or kernel.options.trace_assignment_values:
if codegen_state.vectorization_info and is_vector:
raise Unvectorizable("tracing does not support vectorization")
from cgen import Statement as S # noqa
gs, ls = kernel.get_grid_size_upper_bounds()
printf_format = "%s.%s[%s][%s]: %s" % (kernel.name, insn.id, ", ".join(
"gid%d=%%d" % i for i in range(len(gs))), ", ".join(
"lid%d=%%d" % i for i in range(len(ls))), assignee_var_name)
printf_args = (["gid(%d)" % i for i in range(len(gs))] +
["lid(%d)" % i for i in range(len(ls))])
if assignee_indices:
printf_format += "[%s]" % ",".join(len(assignee_indices) * ["%d"])
printf_args.extend(
ecm(i, prec=PREC_NONE, type_context="i")
for i in assignee_indices)
if kernel.options.trace_assignment_values:
if lhs_dtype.numpy_dtype.kind == "i":
printf_format += " = %d"
printf_args.append(lhs_code)
elif lhs_dtype.numpy_dtype.kind == "f":
printf_format += " = %g"
printf_args.append(lhs_code)
elif lhs_dtype.numpy_dtype.kind == "c":
printf_format += " = %g + %gj"
printf_args.extend(["(%s).x" % lhs_code, "(%s).y" % lhs_code])
if printf_args:
printf_args_str = ", " + ", ".join(printf_args)
else:
printf_args_str = ""
printf_insn = S("printf(\"%s\\n\"%s)" %
(printf_format, printf_args_str))
from cgen import Block
if kernel.options.trace_assignment_values:
result = Block([result, printf_insn])
else:
# print first, execute later -> helps find segfaults
result = Block([printf_insn, result])
# }}}
return result