[pypy-commit] pypy vecopt2: added vector integer field to resop class. removed some specific vec ops (descr might be used for that)
plan_rich
noreply at buildbot.pypy.org
Tue May 5 09:45:49 CEST 2015
Author: Richard Plangger <rich at pasra.at>
Branch: vecopt2
Changeset: r77099:7b76f61e0287
Date: 2015-03-30 17:43 +0200
http://bitbucket.org/pypy/pypy/changeset/7b76f61e0287/
Log: added vector integer field to resop class. removed some specific vec
ops (descr might be used for that) starting to convert packs into
simd instructions refactoring dependecy graph to more efficiently
delete edges
diff --git a/rpython/jit/metainterp/optimizeopt/dependency.py b/rpython/jit/metainterp/optimizeopt/dependency.py
--- a/rpython/jit/metainterp/optimizeopt/dependency.py
+++ b/rpython/jit/metainterp/optimizeopt/dependency.py
@@ -2,9 +2,10 @@
from rpython.jit.metainterp.optimizeopt.util import make_dispatcher_method
from rpython.jit.metainterp.resoperation import rop
from rpython.jit.codewriter.effectinfo import EffectInfo
-from rpython.jit.metainterp.history import BoxPtr, ConstPtr, ConstInt, BoxInt
+from rpython.jit.metainterp.history import BoxPtr, ConstPtr, ConstInt, BoxInt, Box
from rpython.rtyper.lltypesystem import llmemory
from rpython.rlib.unroll import unrolling_iterable
+from rpython.rlib.objectmodel import we_are_translated
MODIFY_COMPLEX_OBJ = [ (rop.SETARRAYITEM_GC, 0, 1)
, (rop.SETARRAYITEM_RAW, 0, 1)
@@ -37,12 +38,6 @@
self.idx_from = idx_from
self.idx_to = idx_to
- def adjust_dep_after_swap(self, idx_old, idx_new):
- if self.idx_from == idx_old:
- self.idx_from = idx_new
- elif self.idx_to == idx_old:
- self.idx_to = idx_new
-
def __repr__(self):
return 'Dep(trace[%d] -> trace[%d], arg: %s)' \
% (self.idx_from, self.idx_to, self.args)
@@ -119,6 +114,8 @@
"""
tracker = DefTracker(self.memory_refs)
+ guards = []
+ # pass 1
for i,op in enumerate(operations):
# the label operation defines all operations at the
# beginning of the loop
@@ -142,14 +139,47 @@
# guard specifics
if op.is_guard():
- for arg in op.getfailargs():
- self._def_use(arg, i, tracker)
- if i > 0:
- self._guard_dependency(op, i, operations, tracker)
+ guards.append(i)
+ # TODO
+ #if i > 0:
+ # self._guard_dependency(op, i, operations, tracker)
+ # pass 2 correct guard dependencies
+ for guard_idx in guards:
+ variables = []
+ for dep in self.depends(guard_idx):
+ idx = dep.idx_from
+ op = operations[idx]
+ for arg in op.getarglist():
+ if isinstance(arg, Box):
+ variables.append(arg)
+ if op.result:
+ variables.append(op.result)
+ print "\ntesting", variables
+ for var in variables:
+ try:
+ def_idx = tracker.definition_index(var)
+ print "guard", guard_idx, def_idx, "var", var, "aaa", [d.idx_to for d in self.get_uses(def_idx)]
+ for dep in self.provides(def_idx):
+ if var in dep.args and dep.idx_to > guard_idx:
+ self._put_edge(guard_idx, dep.idx_to, var)
+ print "put edge", guard_idx, dep.idx_to, var, dep.args
+ except KeyError:
+ pass
+ op = operations[guard_idx]
+ for arg in op.getfailargs():
+ try:
+ def_idx = tracker.definition_index(arg)
+ self._put_edge(def_idx, i, arg)
+ except KeyError:
+ pass
+
+ # pass 3 find schedulable nodes
+ for i,op in enumerate(operations):
if len(self.adjacent_list[i]) == 0:
self.schedulable_nodes.append(i)
+
def update_memory_ref(self, op, index, tracker):
if index not in self.memory_refs:
return
@@ -166,7 +196,7 @@
self.integral_mod.update_memory_ref(memref)
else:
break # an operation that is not tractable
- for dep in self.get_defs(curidx):
+ for dep in self.depends(curidx):
curop = self.operations[dep.idx_from]
if curop.result == memref.origin:
curidx = dep.idx_from
@@ -201,7 +231,7 @@
# A trace is not in SSA form, but this complex object
# modification introduces a WAR/WAW dependency
def_idx = tracker.definition_index(arg)
- for dep in self.get_uses(def_idx):
+ for dep in self.provides(def_idx):
if dep.idx_to >= index:
break
self._put_edge(dep.idx_to, index, argcell)
@@ -226,10 +256,15 @@
if self.modifies_complex_object(op):
for opnum, i, j in unrolling_iterable(MODIFY_COMPLEX_OBJ):
if op.getopnum() == opnum:
+ op_args = op.getarglist()
if j == -1:
args.append((op.getarg(i), None, True))
+ for j in range(i+1,len(op_args)):
+ args.append((op.getarg(j), None, False))
else:
args.append((op.getarg(i), op.getarg(j), True))
+ for x in range(j+1,len(op_args)):
+ args.append((op.getarg(x), None, False))
break
else:
# assume this destroys every argument... can be enhanced by looking
@@ -243,7 +278,7 @@
# respect a guard after a statement that can raise!
assert i > 0
- j = i-1
+ j = i - 1
while j > 0:
prev_op = operations[j]
if prev_op.is_guard():
@@ -275,6 +310,8 @@
def _put_edge(self, idx_from, idx_to, arg):
assert idx_from != idx_to
+ if idx_from == 6 and idx_to == 9:
+ assert False
dep = self.instr_dependency(idx_from, idx_to)
if dep is None:
dep = Dependency(idx_from, idx_to, arg)
@@ -284,13 +321,28 @@
if arg not in dep.args:
dep.args.append(arg)
+ def provides_count(self, idx):
+ i = 0
+ for _ in self.provides(idx):
+ i += 1
+ return i
+
+ def provides(self, idx):
+ return self.get_uses(idx)
def get_uses(self, idx):
for dep in self.adjacent_list[idx]:
if idx < dep.idx_to:
yield dep
+ def depends_count(self, idx):
+ i = 0
+ for _ in self.depends(idx):
+ i += 1
+ return i
+
+ def depends(self, idx):
+ return self.get_defs(idx)
def get_defs(self, idx):
- deps = []
for dep in self.adjacent_list[idx]:
if idx > dep.idx_from:
yield dep
@@ -344,11 +396,25 @@
return edge
return None
+ def remove_depencency(self, follow_dep, point_to_idx):
+ """ removes a all dependencies that point to the second parameter.
+ it is assumed that the adjacent_list[point_to_idx] is not iterated
+ when calling this function.
+ """
+ idx = follow_dep.idx_from
+ if idx == point_to_idx:
+ idx = follow_dep.idx_to
+
+ preount = len(self.adjacent_list[idx])
+ self.adjacent_list[idx] = [d for d in self.adjacent_list[idx] \
+ if d.idx_to != point_to_idx and d.idx_from != point_to_idx]
+ #print "reduced", idx, "from",preount,"to",len(self.adjacent_list[idx])
+
def __repr__(self):
graph = "graph([\n"
for i,l in enumerate(self.adjacent_list):
- graph += " "
+ graph += " " + str(i) + ": "
for d in l:
if i == d.idx_from:
graph += str(d.idx_to) + ","
@@ -358,19 +424,6 @@
return graph + " ])"
- def swap_instructions(self, ia, ib):
- depa = self.adjacent_list[ia]
- depb = self.adjacent_list[ib]
-
- for d in depa:
- d.adjust_dep_after_swap(ia, ib)
-
- for d in depb:
- d.adjust_dep_after_swap(ib, ia)
-
- self.adjacent_list[ia] = depb
- self.adjacent_list[ib] = depa
-
def loads_from_complex_object(self, op):
opnum = op.getopnum()
return rop._ALWAYS_PURE_LAST <= opnum and opnum <= rop._MALLOC_FIRST
@@ -379,6 +432,24 @@
opnum = op.getopnum()
return rop.SETARRAYITEM_GC<= opnum and opnum <= rop.UNICODESETITEM
+ def as_dot(self, operations):
+ if not we_are_translated():
+ dot = "digraph dep_graph {\n"
+
+ for i in range(len(self.adjacent_list)):
+ op = operations[i]
+ dot += " n%d [label=\"[%d]: %s\"];\n" % (i,i,str(op))
+
+ dot += "\n"
+ for i,alist in enumerate(self.adjacent_list):
+ for dep in alist:
+ if dep.idx_to > i:
+ dot += " n%d -> n%d;\n" % (i,dep.idx_to)
+ dot += "\n}\n"
+ return dot
+
+ return ""
+
class Scheduler(object):
def __init__(self, graph):
self.graph = graph
@@ -390,14 +461,49 @@
def next_schedule_index(self):
return self.schedulable_nodes[0]
+ def schedulable(self, indices):
+ for index in indices:
+ if index not in self.schedulable_nodes:
+ break
+ else:
+ return True
+ return False
+
+ def schedule_later(self, index):
+ node = self.schedulable_nodes[index]
+ del self.schedulable_nodes[index]
+ self.schedulable_nodes.append(node)
+ print "shifting", index, "(", node ,")","to", len(self.schedulable_nodes)-1, "sched", self.schedulable_nodes
+
+ def schedule_all(self, opindices):
+ indices = []
+ while len(opindices) > 0:
+ opidx = opindices.pop()
+ for i,node in enumerate(self.schedulable_nodes):
+ if node == opidx:
+ indices.append(i)
+ for index in indices:
+ self.schedule(index)
+
def schedule(self, index):
node = self.schedulable_nodes[index]
del self.schedulable_nodes[index]
+ print "schedule[", index, "](", node, "):",
+ to_del = []
+ adj_list = self.graph.adjacent_list[node]
+ for dep in adj_list:
+ self.graph.remove_depencency(dep, node)
#
- for dep in self.graph.get_uses(node):
- self.schedulable_nodes.append(dep.idx_to)
- #
- # self.graph.adjacent_list[node] = None
+ for dep in self.graph.provideso(node):
+ candidate = dep.idx_to
+ if self.is_schedulable(dep.idx_to):
+ self.schedulable_nodes.append(dep.idx_to)
+ print dep.idx_to, ",",
+ self.graph.adjacent_list[node] = []
+ print ""
+
+ def is_schedulable(self, idx):
+ return self.graph.depends_count(idx) == 0
class IntegralMod(object):
""" Calculates integral modifications on an integer object.
diff --git a/rpython/jit/metainterp/optimizeopt/test/test_dependency.py b/rpython/jit/metainterp/optimizeopt/test/test_dependency.py
--- a/rpython/jit/metainterp/optimizeopt/test/test_dependency.py
+++ b/rpython/jit/metainterp/optimizeopt/test/test_dependency.py
@@ -71,6 +71,14 @@
def assert_dependent(self, a, b):
assert not self.last_graph.independent(a,b), "{a} and {b} are independent!".format(a=a,b=b)
+ def _write_dot_and_convert_to_svg(self, graph, ops, filename):
+ dot = graph.as_dot(ops)
+ with open('/home/rich/' + filename + '.dot', 'w') as fd:
+ fd.write(dot)
+ with open('/home/rich/'+filename+'.svg', 'w') as fd:
+ import subprocess
+ subprocess.Popen(['dot', '-Tsvg', '/home/rich/'+filename+'.dot'], stdout=fd).communicate()
+
class BaseTestDependencyGraph(DepTestHelper):
def test_dependency_empty(self):
ops = """
@@ -130,6 +138,18 @@
self.assert_edges(dep_graph,
[ [2,3], [2], [1,0], [0] ])
+ #def test_dependency_guard_2(self):
+ # ops = """
+ # [i1]
+ # i2 = int_le(i1, 10)
+ # guard_true(i2) [i1]
+ # i3 = int_add(i1,1)
+ # jump(i3)
+ # """
+ # dep_graph = self.build_dependency(ops)
+ # self.assert_edges(dep_graph,
+ # [ [1], [0,2], [1], [2,4], [3] ])
+
def test_no_edge_duplication(self):
ops = """
[i1]
diff --git a/rpython/jit/metainterp/optimizeopt/test/test_vectorize.py b/rpython/jit/metainterp/optimizeopt/test/test_vectorize.py
--- a/rpython/jit/metainterp/optimizeopt/test/test_vectorize.py
+++ b/rpython/jit/metainterp/optimizeopt/test/test_vectorize.py
@@ -83,13 +83,23 @@
def schedule(self, loop, unroll_factor = -1):
opt = self.vec_optimizer_unrolled(loop, unroll_factor)
+ self.debug_print_operations(opt.loop)
opt.build_dependency_graph()
opt.find_adjacent_memory_refs()
+ self._write_dot_and_convert_to_svg(opt.dependency_graph, opt.loop.operations, 'test')
opt.extend_packset()
opt.combine_packset()
opt.schedule()
return opt
+ def _write_dot_and_convert_to_svg(self, graph, ops, filename):
+ dot = graph.as_dot(ops)
+ with open('/home/rich/' + filename + '.dot', 'w') as fd:
+ fd.write(dot)
+ with open('/home/rich/'+filename+'.svg', 'w') as fd:
+ import subprocess
+ subprocess.Popen(['dot', '-Tsvg', '/home/rich/'+filename+'.dot'], stdout=fd).communicate()
+
def assert_unroll_loop_equals(self, loop, expected_loop, \
unroll_factor = -1):
vec_optimizer = self.vec_optimizer_unrolled(loop, unroll_factor)
@@ -107,7 +117,12 @@
def debug_print_operations(self, loop):
print('--- loop instr numbered ---')
for i,op in enumerate(loop.operations):
- print(i,op)
+ print "[",i,"]",op,
+ if op.is_guard():
+ print op.rd_snapshot.boxes
+ else:
+ print ""
+
def assert_pack(self, pack, indices):
assert len(pack.operations) == len(indices)
@@ -833,42 +848,36 @@
self.assert_has_pack_with(vopt.packset, opindices)
def test_schedule_vector_operation(self):
- for op,vop in [('int_add','vec_int_add'), ('int_sub','vec_int_sub'),
- ('int_mul','vec_int_mul')]:
+ for op,vop in [ ('int_mul','vec_int_mul')]: #('int_add','vec_int_add'), ('int_sub','vec_int_sub'),
ops = """
- [p0,p1,p2,i0]
- i1 = int_add(i0, 1)
- i10 = int_le(i1, 128)
- guard_true(i10) []
- i2 = getarrayitem_gc(p0, i0, descr=floatarraydescr)
- i3 = getarrayitem_gc(p1, i0, descr=floatarraydescr)
- i4 = {op}(i2,i3)
- setarrayitem_gc(p2, i0, i4, descr=floatarraydescr)
- jump(p0,p1,p2,i1)
+ [p0,p1,p2,i0] # 0
+ i10 = int_le(i0, 128) # 1, 8, 15, 22
+ guard_true(i10) [p0,p1,p2,i0] # 2, 9, 16, 23
+ i2 = getarrayitem_gc(p0, i0, descr=floatarraydescr) # 3, 10, 17, 24
+ i3 = getarrayitem_gc(p1, i0, descr=floatarraydescr) # 4, 11, 18, 25
+ i4 = {op}(i2,i3) # 5, 12, 19, 26
+ setarrayitem_gc(p2, i0, i4, descr=floatarraydescr) # 6, 13, 20, 27
+ i1 = int_add(i0, 1) # 7, 14, 21, 28
+ jump(p0,p1,p2,i1) # 29
""".format(op=op)
vops = """
[p0,p1,p2,i0]
- i1 = int_add(i0, 1)
i10 = int_le(i1, 128)
guard_true(i10) []
- i11 = int_add(i1, 1)
+ i1 = int_add(i0, 1)
i12 = int_le(i11, 128)
guard_true(i12) []
- i13 = int_add(i11, 1)
- i14 = int_le(i13, 128)
- guard_true(i14) []
- i15 = int_add(i13, 1)
- i16 = int_le(i15, 128)
- guard_true(i16) []
+ i11 = int_add(i1, 1)
i2 = vec_raw_load(p0, i0, 4, descr=floatarraydescr)
i3 = vec_raw_load(p1, i0, 4, descr=floatarraydescr)
i4 = {op}(i2,i3,4,descr=floatarraydescr)
vec_raw_store(p2, i0, i4, 4, descr=floatarraydescr)
- jump(p0,p1,p2,i15)
+ jump(p0,p1,p2,i12)
""".format(op=vop)
loop = self.parse_loop(ops)
- vopt = self.schedule(loop,3)
- self.assert_equal(loop, self.parse_loop(vops))
+ vopt = self.schedule(loop,1)
+ self.debug_print_operations(vopt.loop)
+ #self.assert_equal(loop, self.parse_loop(vops))
class TestLLtype(BaseTestVectorize, LLtypeMixin):
pass
diff --git a/rpython/jit/metainterp/optimizeopt/vectorize.py b/rpython/jit/metainterp/optimizeopt/vectorize.py
--- a/rpython/jit/metainterp/optimizeopt/vectorize.py
+++ b/rpython/jit/metainterp/optimizeopt/vectorize.py
@@ -1,11 +1,12 @@
import sys
import py
from rpython.rtyper.lltypesystem import lltype, rffi
+from rpython.jit.metainterp.history import ConstInt
from rpython.jit.metainterp.optimizeopt.optimizer import Optimizer, Optimization
from rpython.jit.metainterp.optimizeopt.util import make_dispatcher_method
from rpython.jit.metainterp.optimizeopt.dependency import (DependencyGraph,
MemoryRef, IntegralMod, Scheduler)
-from rpython.jit.metainterp.resoperation import rop
+from rpython.jit.metainterp.resoperation import (rop, ResOperation)
from rpython.jit.metainterp.resume import Snapshot
from rpython.rlib.debug import debug_print, debug_start, debug_stop
from rpython.jit.metainterp.jitexc import JitException
@@ -40,6 +41,7 @@
self.unroll_count = 0
def emit_operation(self, op):
+ print "emit[", len(self._newoperations), "]:", op
self._last_emitted_op = op
self._newoperations.append(op)
@@ -269,6 +271,7 @@
def combine_packset(self):
if len(self.packset.packs) == 0:
raise NotAVectorizeableLoop()
+ # TODO modifying of lists while iterating has undefined results!!
while True:
len_before = len(self.packset.packs)
for i,pack1 in enumerate(self.packset.packs):
@@ -285,7 +288,9 @@
break
def schedule(self):
+ self.clear_newoperations()
scheduler = Scheduler(self.dependency_graph)
+ i = 0
while scheduler.has_more_to_schedule():
candidate_index = scheduler.next_schedule_index()
candidate = self.loop.operations[candidate_index]
@@ -295,13 +300,32 @@
else:
self.emit_operation(candidate)
scheduler.schedule(0)
+ i+=1
+ if i > 20:
+ print self.dependency_graph
+ break
+
+ self.loop.operations = self._newoperations[:]
def _schedule_pack(self, scheduler, pack):
- if scheduler.all_schedulable([ e.opidx for e in pack.operations ]):
+ opindices = [ e.opidx for e in pack.operations ]
+ if scheduler.schedulable(opindices):
self.emit_vec_operation(pack)
+ scheduler.schedule_all(opindices)
+ else:
+ print "pack not schedulable", pack
+ scheduler.schedule_later(0)
def emit_vec_operation(self, pack):
- pass
+ op0_wrapper = pack.operations[0]
+ op0 = self.loop.operations[op0_wrapper.opidx]
+ op_count = len(pack.operations)
+ assert op0.vector != -1
+ args = op0.getarglist()[:]
+ args.append(ConstInt(op_count))
+ vecop = ResOperation(op0.vector, args, op0.result, op0.getdescr())
+ self.emit_operation(vecop)
+
def isomorphic(l_op, r_op):
""" Described in the paper ``Instruction-Isomorphism in Program Execution''.
@@ -364,6 +388,7 @@
return 0
def combine(self, i, j):
+ # TODO modifying of lists while iterating has undefined results!!
pack_i = self.packs[i]
pack_j = self.packs[j]
operations = pack_i.operations
@@ -380,6 +405,13 @@
self.packs[j] = self.packs[last_pos]
del self.packs[last_pos]
+ def pack_for_operation(self, op, opidx):
+ for pack in self.packs:
+ for op in pack.operations:
+ if op.getopidx() == opidx:
+ return pack
+ return None
+
class Pack(object):
""" A pack is a set of n statements that are:
* isomorphic
@@ -417,6 +449,9 @@
self.opidx = opidx
self.memref = memref
+ def getopidx(self):
+ return self.opidx
+
def __eq__(self, other):
if isinstance(other, PackOpWrapper):
return self.opidx == other.opidx and self.memref == other.memref
diff --git a/rpython/jit/metainterp/resoperation.py b/rpython/jit/metainterp/resoperation.py
--- a/rpython/jit/metainterp/resoperation.py
+++ b/rpython/jit/metainterp/resoperation.py
@@ -25,6 +25,7 @@
_cls_has_bool_result = False
boolreflex = -1
boolinverse = -1
+ vector = -1
_attrs_ = ('result',)
@@ -445,21 +446,9 @@
#
# vector operations
'_VEC_ARITHMETIC_FIRST',
- 'VEC_CHAR_ADD/3d',
- 'VEC_CHAR_SUB/3d',
- 'VEC_CHAR_MUL/3d',
- 'VEC_SHORT_ADD/3d',
- 'VEC_SHORT_SUB/3d',
- 'VEC_SHORT_MUL/3d',
'VEC_INT_ADD/3d',
'VEC_INT_SUB/3d',
'VEC_INT_MUL/3d',
- 'VEC_UINT_ADD/3d',
- 'VEC_UINT_SUB/3d',
- 'VEC_UINT_MUL/3d',
- 'VEC_SP_FLOAT_ADD/3d',
- 'VEC_SP_FLOAT_SUB/3d',
- 'VEC_SP_FLOAT_MUL/3d',
'VEC_FLOAT_ADD/3d',
'VEC_FLOAT_SUB/3d',
'VEC_FLOAT_MUL/3d',
@@ -707,6 +696,22 @@
rop.PTR_EQ: rop.PTR_EQ,
rop.PTR_NE: rop.PTR_NE,
}
+_opvector = {
+ rop.RAW_LOAD: rop.VEC_RAW_LOAD,
+ rop.GETARRAYITEM_RAW: rop.VEC_RAW_LOAD,
+ rop.GETARRAYITEM_GC: rop.VEC_RAW_LOAD,
+
+ rop.RAW_STORE: rop.VEC_RAW_STORE,
+ rop.SETARRAYITEM_RAW: rop.VEC_RAW_STORE,
+ rop.SETARRAYITEM_GC: rop.VEC_RAW_STORE,
+
+ rop.INT_ADD: rop.VEC_INT_ADD,
+ rop.INT_SUB: rop.VEC_INT_SUB,
+ rop.INT_MUL: rop.VEC_INT_MUL,
+ rop.FLOAT_ADD: rop.VEC_FLOAT_ADD,
+ rop.FLOAT_SUB: rop.VEC_FLOAT_SUB,
+ rop.FLOAT_MUL: rop.VEC_FLOAT_MUL,
+}
def setup2():
for cls in opclasses:
@@ -717,10 +722,13 @@
cls.boolreflex = _opboolreflex[opnum]
if opnum in _opboolinverse:
cls.boolinverse = _opboolinverse[opnum]
+ if opnum in _opvector:
+ cls.vector = _opvector[opnum]
setup2()
del _opboolinverse
del _opboolreflex
+del _opvector
def get_deep_immutable_oplist(operations):
"""
More information about the pypy-commit
mailing list