From 89deff324639aab9cd28d6f47c29259d02e3a036 Mon Sep 17 00:00:00 2001
From: Andreas Kloeckner <inform@tiker.net>
Date: Mon, 24 Sep 2012 18:05:02 -0400
Subject: [PATCH] In CSE: Only inherit restrictions unrelated to storage from
global s2s map.
---
MEMO | 10 +++++-----
loopy/cse.py | 11 +++++------
2 files changed, 10 insertions(+), 11 deletions(-)
diff --git a/MEMO b/MEMO
index 23d230940..3087e53bb 100644
--- a/MEMO
+++ b/MEMO
@@ -16,9 +16,6 @@ Things to consider
- Every loop in loopy is opened at most once.
Too restrictive?
-- Loop bounds currently may not depend on parallel dimensions
- Does it make sense to relax this?
-
- Why do CSEs necessarily have to duplicate the inames?
-> because that would be necessary for a sequential prefetch
@@ -49,8 +46,6 @@ To-do
- Fix timer / call code
-- make sure simple side effects in global work
-
- Fuse: store/fetch elimination?
Fixes:
@@ -115,6 +110,11 @@ Future ideas
Dealt with
^^^^^^^^^^
+- make sure simple side effects work
+
+- Loop bounds currently may not depend on parallel dimensions
+ Does it make sense to relax this?
+
- Streamline argument specification
- syntax for linear array access
diff --git a/loopy/cse.py b/loopy/cse.py
index 71b7b3c46..bd5ca733a 100644
--- a/loopy/cse.py
+++ b/loopy/cse.py
@@ -197,13 +197,12 @@ def build_global_storage_to_sweep_map(kernel, invocation_descriptors,
s2s_domain, aligned_g_s2s_parm_dom = isl.align_two(
s2s_domain, global_s2s_par_dom)
- s2s_domain = s2s_domain.project_out_except(
- arg_inames, [dim_type.param])
- aligned_g_s2s_parm_dom = aligned_g_s2s_parm_dom.project_out_except(
- arg_inames, [dim_type.param])
+ arg_restrictions = (
+ aligned_g_s2s_parm_dom
+ .eliminate(dim_type.set, 0, aligned_g_s2s_parm_dom.dim(dim_type.set))
+ .remove_divs())
- is_in_footprint = ((s2s_domain & aligned_g_s2s_parm_dom)
- .is_subset(aligned_g_s2s_parm_dom))
+ is_in_footprint = (arg_restrictions & s2s_domain).is_subset(aligned_g_s2s_parm_dom)
invdesc.is_in_footprint = is_in_footprint
--
GitLab