From ebad5755515c7ce7a721c05fd7e0d1f956304ad8 Mon Sep 17 00:00:00 2001
From: Marcel Koch <marcel.koch@uni-muenster.de>
Date: Fri, 26 Jan 2018 15:11:56 +0100
Subject: [PATCH] fix cherry picks
---
.../perftool/blockstructured/vectorization.py | 65 +++++++++----------
1 file changed, 31 insertions(+), 34 deletions(-)
diff --git a/python/dune/perftool/blockstructured/vectorization.py b/python/dune/perftool/blockstructured/vectorization.py
index 6eba316b..18f8fb02 100644
--- a/python/dune/perftool/blockstructured/vectorization.py
+++ b/python/dune/perftool/blockstructured/vectorization.py
@@ -49,14 +49,14 @@ def add_vcl_vector(knl, iname_inner, iname_outer):
new_vec_temporaries[vec] = DuneTemporaryVariable(vec, dtype=np.float64, shape=(4,), managed=True,
scope=lp.temp_var_scope.PRIVATE, dim_tags=('vec',))
- # write accum expr as "expr + r"
modified_accum_insn = []
replace_accum_insn = dict()
vng = knl.get_var_name_generator()
idg = knl.get_instruction_id_generator()
for insn in write_insns:
if isinstance(insn, lp.Assignment):
- expr_without_r = prim.Sum(tuple(e for e in insn.expression.children if not e ==insn.assignee))
+ # write accum expr as "r = expr + r"
+ expr_without_r = prim.Sum(tuple(e for e in insn.expression.children if not e == insn.assignee))
if expr_without_r == insn.expression:
continue
# finde micro inames
@@ -109,12 +109,11 @@ def add_vcl_vector(knl, iname_inner, iname_outer):
# r+=a[iy]
id_accum = idg('insn_mod_accum')
expr_accum = prim.Sum((var_a, prim.Call(prim.Variable('permute4d<-1,0,1,2>'), (var_b,)),
- substitute(insn.assignee,{iname_ix:0})))
+ substitute(insn.assignee, {iname_ix:0})))
replace_accum_insn[insn.id] = lp.Assignment(assignee=substitute(insn.assignee,{iname_ix:0}),
expression=expr_accum,
- id='insn_mod_accum',
- depends_on=insn.depends_on|frozenset({'insn_b', 'insn_init_a',
- 'insn_a_iy'}),
+ id=id_accum,
+ depends_on=insn.depends_on|frozenset({id_set_a,id_init_a,id_set_b}),
within_inames=insn.within_inames-frozenset({iname_ix})
)
# a[iy] = permute
@@ -189,41 +188,39 @@ def add_vcl_vector(knl, iname_inner, iname_outer):
# flat index without vec iname
strides = tuple(tag.stride for tag in knl.temporary_variables[alias].dim_tags)
- index = prim.Sum(tuple(prim.Product(z) for z in zip(substitute(expr, {iname_inner:0}).index_tuple, strides)))
-
- # add store instruction
- code = "{}.store({}+ {});".format(vec, alias, index)
- #store_id = knl.make_unique_instruction_id(insns=knl.instructions+load_insns, based_on='insn_'+vec+'_store')
- store_id = idg('insn_'+vec+'_store')
- store_insns.append(lp.CInstruction(iname_exprs=[], code=code,read_variables=frozenset({alias}),
- within_inames=insn.within_inames,
- depends_on=insn.depends_on
- | frozenset({insn.id})
- | read_dependencies[insn.id],
- id=store_id))
+ index = prim.Sum(tuple(prim.Product(z) for z in zip(substitute(expr, {iname_inner: 0, iname_ix: 0}).index_tuple, strides)))
+
+ # add store instruction
+ code = "{}.store({} + {});".format(vec, alias, index)
+ #store_id = knl.make_unique_instruction_id(insns=knl.instructions+load_insns, based_on='insn_'+vec+'_store')
+ store_id = idg('insn_'+vec+'_store')
+ store_insns.append(lp.CInstruction(iname_exprs=[], code=code,read_variables=frozenset({alias}),
+ within_inames=insn.within_inames,
+ depends_on=insn.depends_on
+ | frozenset({insn.id})
+ | read_dependencies[insn.id],
+ id=store_id))
# exchange alias for vector
new_insns = []
for insn in knl.instructions:
+ insn = replace_accum_insn.get(insn.id, insn)
if insn.id not in read_insn_to_vec_instance.keys() | write_insn_to_vec_instance.keys():
new_insns.append(insn)
else:
- if insn.id in replace_accum_insn:
- new_insn = replace_accum_insn[insn.id].copy(depends_on=replace_accum_insn[insn.id].depends_on
- | read_dependencies[insn.id])
- else:
- subst_map = dict()
- for vec, expr in read_insn_to_vec_instance[insn.id]:
- subst_map[expr] = prim.Subscript(prim.Variable(vec), (prim.Variable(iname_inner),))
-
- new_insn = insn
-
- if insn in read_insns:
- new_insn = new_insn.copy(expression=substitute(new_insn.expression, subst_map),
- depends_on=new_insn.depends_on | read_dependencies[insn.id])
- if insn in write_insns:
- new_insn = new_insn.copy(assignee=substitute(new_insn.assignee, subst_map))
- new_insns.append(new_insn)
+ subst_map = dict()
+ for vec, expr in read_insn_to_vec_instance[insn.id]:
+ subst_map[expr] = prim.Subscript(prim.Variable(vec), (prim.Variable(iname_inner),))
+
+ new_insn = insn
+
+ if insn in read_insns:
+ new_insn = new_insn.copy(expression=substitute(new_insn.expression, subst_map),
+ depends_on=new_insn.depends_on | read_dependencies[insn.id])
+ if insn in write_insns:
+ new_insn = new_insn.copy(assignee=substitute(new_insn.assignee, subst_map))
+
+ new_insns.append(new_insn)
from loopy.kernel.creation import resolve_dependencies
return resolve_dependencies(knl.copy(instructions=new_insns+load_insns+store_insns+modified_accum_insn,
--
GitLab