Actions
 |
Post |
 |
Subscribe |
|
Unsubscribe |
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[patch] Parallel support function updates
- To: VSIPL++ Developers List <vsipl++@xxxxxxxxxxxxxxxx>
- Subject: [patch] Parallel support function updates
- From: Jules Bergmann <jules@xxxxxxxxxxxxxxxx>
- Date: Fri, 24 Mar 2006 13:58:29 -0500
This patch updates the parallel support functions provided by the library to match what is covered in the specification. This includes both new functions, and extending some existing functions to work with more distribution types (in particular, the global to local index conversions now work with cyclic distributions). It cleans up the map's internal parallel support API to be more consistent.
It implements the Replicated_map map, which is similar to the existing Global_map but allows replication over a subset of processors.
This patch fixes a bug with creating local views of sliced subviews of distributed objects (subviews using Sliced_block), and adds a regression test. In particular, if 'distributed_matrix' was a distributed matrix, the local view of a row or column subview:
distributed_matrix.row(0).local() would be invalid on all other processors than the one(s) owning row(0). -- Jules -- Jules Bergmann CodeSourcery jules@xxxxxxxxxxxxxxxx (650) 331-3385 x705
Index: ChangeLog
===================================================================
RCS file: /home/cvs/Repository/vpp/ChangeLog,v
retrieving revision 1.420
diff -u -r1.420 ChangeLog
--- ChangeLog 24 Mar 2006 12:36:05 -0000 1.420
+++ ChangeLog 24 Mar 2006 18:55:40 -0000
@@ -1,5 +1,62 @@
2006-03-24 Jules Bergmann <jules@xxxxxxxxxxxxxxxx>
+ * src/vsip/dense.hpp: Rename implementation par support functions
+ with impl_ prefix.
+ * src/vsip/impl/distributed-block.hpp: Likewise.
+ * src/vsip/map.hpp: Rename implementation par support functions
+ with impl_ prefix.
+ (impl_patch_from_global_index, impl_dim_patch_from_index): New
+ functions to compute patch from global index.
+ (Is_global_map): Specialize trait for Map class.
+ * src/vsip/parallel.hpp: Include par-support.hpp.
+ * src/vsip/support.hpp (no_index): New constant value.
+ * src/vsip/impl/block-copy.hpp: Fix handling of distributed blocks.
+ * src/vsip/impl/block-traits.hpp (Is_par_same_map): Describe what
+ it means to be "same map".
+ * src/vsip/impl/dispatch-assign.hpp: Move map traits into
+ map-traits.hpp.
+ * src/vsip/impl/dist.hpp: Consistent names for impl par support
+ functions.
+ (impl_subblock_from_index): Implement function for Cyclic_dist.
+ (impl_local_from_global_index): Likewise.
+ * src/vsip/impl/global_map.hpp: Rename implementation par support
+ functions with impl_ prefix. Specialize map traits.
+ * src/vsip/impl/local_map.hpp: Likewise.
+ * src/vsip/impl/layout.hpp: Add assertions that indices are valid.
+ * src/vsip/impl/map-traits.hpp (Is_global_map, Is_local_only,
+ Is_global_only): Define general traits for maps.
+ * src/vsip/impl/par-assign.hpp: Update use of implementation par
+ support functions. Remove unnecessary assertions.
+ * src/vsip/impl/par-chain-assign.hpp: Likewise.
+ * src/vsip/impl/par-util.hpp: Move parallel support functions
+ [view.support.fcn] to ...
+ * src/vsip/impl/par-util.hpp: New file, ... here.
+ * src/vsip/impl/replicated_map.hpp: New file, implement
+ Replicated_map class.
+ * src/vsip/impl/subblock.hpp: Rename implementation par support
+ functions with impl_ prefix.
+ (Sliced_block): Recognize special index 'no_index' to indicate
+ empty slice. Use this for get_local_block when local view is
+ empty.
+ * src/vsip/impl/sv_block.hpp: New file, block with std::vector as
+ storage. Intended for distributed implementation use only.
+ * tests/appmap.cpp: Update test coverage for parallel support
+ functions, including impl_{subblock,patch}_from_global_index.
+ * tests/replicated_map.cpp: New file, unit tests for Replicated_map.
+ * tests/util-par.hpp: Update use of parallel support functions.
+ * tests/parallel/block.cpp: Likewise.
+ * tests/parallel/expr.cpp: Likewise.
+ * tests/parallel/fftm.cpp: Likewise.
+ * tests/parallel/subviews.cpp: Likewise.
+ * tests/parallel/user-storage.cpp: Likewise.
+ * tests/parallel/corner-turn.cpp: New test, simple corner-turn.
+ * tests/parallel/replicated_data.cpp: New file, unit tests for
+ views distributed with Replicated_map.
+ * tests/regressions/localview_of_slice.cpp: New file, regression
+ test for local view of a sliced block.
+
+2006-03-24 Jules Bergmann <jules@xxxxxxxxxxxxxxxx>
+
* benchmarks/loop.hpp: Fix macro typo. Fix Wall warnings.
Use different loop variables for nested loops.
Index: src/vsip/dense.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/dense.hpp,v
retrieving revision 1.34
diff -u -r1.34 dense.hpp
--- src/vsip/dense.hpp 7 Mar 2006 02:15:22 -0000 1.34
+++ src/vsip/dense.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005,2006 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/dense.hpp
@author Jules Bergmann
@@ -1261,7 +1261,7 @@
map_ (map),
admitted_ (true)
{
- map_.apply(dom);
+ map_.impl_apply(dom);
}
@@ -1281,7 +1281,7 @@
map_ (map),
admitted_ (true)
{
- map_.apply(dom);
+ map_.impl_apply(dom);
}
Index: src/vsip/map.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/map.hpp,v
retrieving revision 1.21
diff -u -r1.21 map.hpp
--- src/vsip/map.hpp 11 Jan 2006 16:22:43 -0000 1.21
+++ src/vsip/map.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/map.hpp
@author Jules Bergmann
@@ -27,6 +27,7 @@
#include <vsip/impl/block-traits.hpp>
#include <vsip/impl/global_map.hpp>
+#include <vsip/impl/replicated_map.hpp>
@@ -153,20 +154,20 @@
// Applied map functions.
- length_type num_patches (index_type sb) const VSIP_NOTHROW;
+ length_type impl_num_patches (index_type sb) const VSIP_NOTHROW;
template <dimension_type Dim>
- void apply(Domain<Dim> const& dom) VSIP_NOTHROW;
+ void impl_apply(Domain<Dim> const& dom) VSIP_NOTHROW;
template <dimension_type Dim>
- Domain<Dim> subblock_domain(index_type sb) const VSIP_NOTHROW;
+ Domain<Dim> impl_subblock_domain(index_type sb) const VSIP_NOTHROW;
template <dimension_type Dim>
- Domain<Dim> global_domain(index_type sb, index_type patch)
+ Domain<Dim> impl_global_domain(index_type sb, index_type patch)
const VSIP_NOTHROW;
template <dimension_type Dim>
- Domain<Dim> local_domain (index_type sb, index_type patch)
+ Domain<Dim> impl_local_domain (index_type sb, index_type patch)
const VSIP_NOTHROW;
template <dimension_type Dim>
@@ -175,17 +176,14 @@
// Implementation functions.
impl_pvec_type impl_pvec() const { return pvec_; }
impl::Communicator impl_comm() const { return comm_; }
- processor_type impl_rank() const { return comm_.rank(); }
- length_type impl_size() const { return comm_.size(); }
bool impl_is_applied() const { return dim_ != 0; }
length_type impl_working_size() const
{ return std::min(this->num_subblocks(), this->pvec_.size()); }
-
- // Private implementation functions.
-private:
+ // Implementation functions.
+public:
length_type impl_subblock_patches(dimension_type d, index_type sb)
const VSIP_NOTHROW;
length_type impl_subblock_size(dimension_type d, index_type sb)
@@ -196,8 +194,10 @@
Domain<1> impl_patch_local_dom(dimension_type d, index_type sb,
index_type p)
const VSIP_NOTHROW;
-public:
- index_type impl_subblock_from_index(dimension_type d, index_type idx)
+
+ index_type impl_dim_subblock_from_index(dimension_type d, index_type idx)
+ const VSIP_NOTHROW;
+ index_type impl_dim_patch_from_index(dimension_type d, index_type idx)
const VSIP_NOTHROW;
index_type impl_local_from_global_index(dimension_type d, index_type idx)
const VSIP_NOTHROW;
@@ -206,8 +206,12 @@
index_type impl_subblock_from_global_index(Index<Dim> const& idx)
const VSIP_NOTHROW;
- index_type global_from_local_index(dimension_type d, index_type sb,
- index_type idx)
+ template <dimension_type Dim>
+ index_type impl_patch_from_global_index(Index<Dim> const& idx)
+ const VSIP_NOTHROW;
+
+ index_type impl_global_from_local_index(dimension_type d, index_type sb,
+ index_type idx)
const VSIP_NOTHROW;
template <dimension_type Dim>
@@ -225,9 +229,9 @@
friend bool map_equiv<>(Map const&, Map const&) VSIP_NOTHROW;
public:
- index_type lookup_index(processor_type pr) const;
- processor_type impl_proc(index_type idx) const { return pvec_[idx]; }
- length_type impl_num_procs() const { return num_procs_; }
+ index_type impl_rank_from_proc(processor_type pr) const;
+ processor_type impl_proc_from_rank(index_type idx) const
+ { return pvec_[idx]; }
// Members.
private:
@@ -373,7 +377,7 @@
typename Dist2>
template <dimension_type Dim>
inline void
-Map<Dist0, Dist1, Dist2>::apply(Domain<Dim> const& dom)
+Map<Dist0, Dist1, Dist2>::impl_apply(Domain<Dim> const& dom)
VSIP_NOTHROW
{
Domain<1> arr[VSIP_MAX_DIMENSION];
@@ -557,7 +561,7 @@
typename Dist1,
typename Dist2>
inline index_type
-Map<Dist0, Dist1, Dist2>::impl_subblock_from_index(
+Map<Dist0, Dist1, Dist2>::impl_dim_subblock_from_index(
dimension_type d,
index_type idx
)
@@ -577,6 +581,32 @@
+// Get local patch for a given index.
+
+template <typename Dist0,
+ typename Dist1,
+ typename Dist2>
+inline index_type
+Map<Dist0, Dist1, Dist2>::impl_dim_patch_from_index(
+ dimension_type d,
+ index_type idx
+ )
+ const VSIP_NOTHROW
+{
+ assert(d < VSIP_MAX_DIMENSION);
+ assert(d < dim_);
+
+ switch (d)
+ {
+ default: assert(false);
+ case 0: return dist0_.impl_patch_from_index(dom_[0], idx);
+ case 1: return dist1_.impl_patch_from_index(dom_[1], idx);
+ case 2: return dist2_.impl_patch_from_index(dom_[2], idx);
+ }
+}
+
+
+
template <typename Dist0,
typename Dist1,
typename Dist2>
@@ -621,7 +651,7 @@
assert(idx[d] < dom_[d].size());
if (d != 0)
sb *= subblocks_[d];
- sb += impl_subblock_from_index(d, idx[d]);
+ sb += impl_dim_subblock_from_index(d, idx[d]);
}
assert(sb < num_subblocks_);
@@ -629,7 +659,7 @@
impl::split_tuple(sb, dim_, subblocks_, dim_sb);
for (dimension_type d=0; d<Dim; ++d)
{
- assert(dim_sb[d] == impl_subblock_from_index(d, idx[d]));
+ assert(dim_sb[d] == impl_dim_subblock_from_index(d, idx[d]));
}
return sb;
@@ -637,6 +667,40 @@
+/// Determine subblock/patch holding a global index.
+
+template <typename Dist0,
+ typename Dist1,
+ typename Dist2>
+template <dimension_type Dim>
+inline index_type
+Map<Dist0, Dist1, Dist2>::impl_patch_from_global_index(
+ Index<Dim> const& idx
+ )
+ const VSIP_NOTHROW
+{
+ index_type p = 0;
+ index_type dim_sb[VSIP_MAX_DIMENSION];
+
+ assert(dim_ != 0 && dim_ == Dim);
+
+ index_type sb = this->impl_subblock_from_global_index(idx);
+ impl::split_tuple(sb, dim_, subblocks_, dim_sb);
+
+ for (dimension_type d=0; d<Dim; ++d)
+ {
+ assert(idx[d] < dom_[d].size());
+ if (d != 0)
+ p *= impl_subblock_patches(d, dim_sb[d]);
+ p += impl_dim_patch_from_index(d, idx[d]);
+ }
+
+ assert(p < this->impl_num_patches(sb));
+ return p;
+}
+
+
+
/// Determine global index from local index for a single dimension
/// Requires:
@@ -650,7 +714,7 @@
typename Dist1,
typename Dist2>
inline index_type
-Map<Dist0, Dist1, Dist2>::global_from_local_index(
+Map<Dist0, Dist1, Dist2>::impl_global_from_local_index(
dimension_type d,
index_type sb,
index_type idx
@@ -731,7 +795,7 @@
typename Dist1,
typename Dist2>
inline index_type
-Map<Dist0, Dist1, Dist2>::lookup_index(processor_type pr)
+Map<Dist0, Dist1, Dist2>::impl_rank_from_proc(processor_type pr)
const
{
for (index_type i=0; i<pvec_.size(); ++i)
@@ -758,7 +822,7 @@
Map<Dist0, Dist1, Dist2>::subblock(processor_type pr)
const VSIP_NOTHROW
{
- index_type pi = lookup_index(pr);
+ index_type pi = impl_rank_from_proc(pr);
if (pi != no_processor && pi < num_subblocks_)
return pi;
@@ -781,8 +845,8 @@
Map<Dist0, Dist1, Dist2>::subblock()
const VSIP_NOTHROW
{
- processor_type pr = impl_rank();
- index_type pi = lookup_index(pr);
+ processor_type pr = local_processor();
+ index_type pi = impl_rank_from_proc(pr);
if (pi != no_processor && pi < num_subblocks_)
return pi;
@@ -879,7 +943,7 @@
typename Dist2>
inline
length_type
-Map<Dist0, Dist1, Dist2>::num_patches(index_type sb)
+Map<Dist0, Dist1, Dist2>::impl_num_patches(index_type sb)
const VSIP_NOTHROW
{
assert(sb < num_subblocks_);
@@ -909,7 +973,7 @@
template <dimension_type Dim>
inline
Domain<Dim>
-Map<Dist0, Dist1, Dist2>::subblock_domain(index_type sb)
+Map<Dist0, Dist1, Dist2>::impl_subblock_domain(index_type sb)
const VSIP_NOTHROW
{
assert(sb < num_subblocks_);
@@ -940,7 +1004,7 @@
template <dimension_type Dim>
inline
Domain<Dim>
-Map<Dist0, Dist1, Dist2>::global_domain(
+Map<Dist0, Dist1, Dist2>::impl_global_domain(
index_type sb,
index_type p)
const VSIP_NOTHROW
@@ -960,7 +1024,7 @@
}
assert(sb < num_subblocks_);
- assert(p < num_patches(sb));
+ assert(p < this->impl_num_patches(sb));
assert(dim_ == Dim);
impl::split_tuple(sb, Dim, subblocks_, dim_sb);
@@ -990,7 +1054,7 @@
template <dimension_type Dim>
inline
Domain<Dim>
-Map<Dist0, Dist1, Dist2>::local_domain(
+Map<Dist0, Dist1, Dist2>::impl_local_domain(
index_type sb,
index_type p
)
@@ -1011,7 +1075,7 @@
}
assert(sb < num_subblocks_);
- assert(p < num_patches(sb));
+ assert(p < this->impl_num_patches(sb));
assert(dim_ == Dim);
@@ -1136,6 +1200,14 @@
namespace impl
{
+template <typename Dist0,
+ typename Dist1,
+ typename Dist2>
+struct Is_global_map<Map<Dist0, Dist1, Dist2> >
+{ static bool const value = true; };
+
+
+
template <typename Dim0,
typename Dim1,
typename Dim2>
@@ -1203,12 +1275,12 @@
length_type num_sb_1 = map.num_subblocks(1);
length_type num_sb_2 = map.num_subblocks(2);
- index_type fix_sb_0 = map.impl_subblock_from_index(0, idx);
+ index_type fix_sb_0 = map.impl_dim_subblock_from_index(0, idx);
Vector<processor_type> pvec(num_sb_1*num_sb_2);
for (index_type pi=0; pi<num_sb_1*num_sb_2; ++pi)
- pvec(pi) = map.impl_proc(fix_sb_0*num_sb_1*num_sb_2+pi);
+ pvec(pi) = map.impl_proc_from_rank(fix_sb_0*num_sb_1*num_sb_2+pi);
return type(pvec, copy_dist<Dist1>(map, 1), copy_dist<Dist2>(map, 2));
}
@@ -1229,7 +1301,7 @@
length_type num_sb_1 = map.num_subblocks(1);
length_type num_sb_2 = map.num_subblocks(2);
- index_type fix_sb_1 = map.impl_subblock_from_index(1, idx);
+ index_type fix_sb_1 = map.impl_dim_subblock_from_index(1, idx);
Vector<processor_type> pvec(num_sb_0*num_sb_2);
@@ -1237,7 +1309,8 @@
{
index_type sb_0 = pi / num_sb_2;
index_type sb_2 = pi % num_sb_2;
- pvec(pi) = map.impl_proc(sb_0*num_sb_1*num_sb_2+fix_sb_1*num_sb_2+sb_2);
+ pvec(pi) = map.impl_proc_from_rank(sb_0*num_sb_1*num_sb_2 +
+ fix_sb_1*num_sb_2 + sb_2);
}
return type(pvec, copy_dist<Dist1>(map, 0), copy_dist<Dist2>(map, 2));
@@ -1259,7 +1332,7 @@
length_type num_sb_1 = map.num_subblocks(1);
length_type num_sb_2 = map.num_subblocks(2);
- index_type fix_sb_2 = map.impl_subblock_from_index(2, idx);
+ index_type fix_sb_2 = map.impl_dim_subblock_from_index(2, idx);
Vector<processor_type> pvec(num_sb_0*num_sb_1);
@@ -1267,9 +1340,9 @@
{
index_type sb_0 = pi / num_sb_1;
index_type sb_1 = pi % num_sb_1;
- pvec(pi) = map.impl_proc(sb_0*num_sb_1*num_sb_2 +
- sb_1*num_sb_2 +
- fix_sb_2);
+ pvec(pi) = map.impl_proc_from_rank(sb_0*num_sb_1*num_sb_2 +
+ sb_1*num_sb_2 +
+ fix_sb_2);
}
return type(pvec, copy_dist<Dist1>(map, 0), copy_dist<Dist2>(map, 1));
@@ -1298,14 +1371,14 @@
length_type num_sb_1 = map.num_subblocks(1);
length_type num_sb_2 = map.num_subblocks(2);
- index_type fix_sb_0 = map.impl_subblock_from_index(0, idx0);
- index_type fix_sb_1 = map.impl_subblock_from_index(1, idx1);
+ index_type fix_sb_0 = map.impl_dim_subblock_from_index(0, idx0);
+ index_type fix_sb_1 = map.impl_dim_subblock_from_index(1, idx1);
Vector<processor_type> pvec(num_sb_2);
for (index_type pi=0; pi<num_sb_2; ++pi)
- pvec(pi) = map.impl_proc(fix_sb_0*num_sb_1*num_sb_2 +
- fix_sb_1*num_sb_2 + pi);
+ pvec(pi) = map.impl_proc_from_rank(fix_sb_0*num_sb_1*num_sb_2 +
+ fix_sb_1*num_sb_2 + pi);
return type(pvec, copy_dist<Dist2>(map, 2));
}
@@ -1327,14 +1400,14 @@
length_type num_sb_1 = map.num_subblocks(1);
length_type num_sb_2 = map.num_subblocks(2);
- index_type fix_sb_0 = map.impl_subblock_from_index(0, idx0);
- index_type fix_sb_2 = map.impl_subblock_from_index(2, idx2);
+ index_type fix_sb_0 = map.impl_dim_subblock_from_index(0, idx0);
+ index_type fix_sb_2 = map.impl_dim_subblock_from_index(2, idx2);
Vector<processor_type> pvec(num_sb_1);
for (index_type pi=0; pi<num_sb_1; ++pi)
- pvec(pi) = map.impl_proc(fix_sb_0*num_sb_1*num_sb_2 +
- pi*num_sb_2 + fix_sb_2);
+ pvec(pi) = map.impl_proc_from_rank(fix_sb_0*num_sb_1*num_sb_2 +
+ pi*num_sb_2 + fix_sb_2);
return type(pvec, copy_dist<Dist1>(map, 1));
}
@@ -1356,14 +1429,14 @@
length_type num_sb_1 = map.num_subblocks(1);
length_type num_sb_2 = map.num_subblocks(2);
- index_type fix_sb_1 = map.impl_subblock_from_index(1, idx1);
- index_type fix_sb_2 = map.impl_subblock_from_index(2, idx2);
+ index_type fix_sb_1 = map.impl_dim_subblock_from_index(1, idx1);
+ index_type fix_sb_2 = map.impl_dim_subblock_from_index(2, idx2);
Vector<processor_type> pvec(num_sb_0);
for (index_type pi=0; pi<num_sb_0; ++pi)
- pvec(pi) = map.impl_proc(pi*num_sb_1*num_sb_2 +
- fix_sb_1*num_sb_2 + fix_sb_2);
+ pvec(pi) = map.impl_proc_from_rank(pi*num_sb_1*num_sb_2 +
+ fix_sb_1*num_sb_2 + fix_sb_2);
return type(pvec, copy_dist<Dist0>(map, 0));
}
@@ -1407,7 +1480,7 @@
Vector<processor_type> pvec(map.num_subblocks());
for (index_type pi=0; pi<map.num_subblocks(); ++pi)
- pvec(pi) = map.impl_proc(pi);
+ pvec(pi) = map.impl_proc_from_rank(pi);
return type(pvec,
copy_dist<Dist0>(map, 0),
Index: src/vsip/parallel.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/parallel.hpp,v
retrieving revision 1.4
diff -u -r1.4 parallel.hpp
--- src/vsip/parallel.hpp 11 Jan 2006 16:22:44 -0000 1.4
+++ src/vsip/parallel.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/parallel.hpp
@author Jules Bergmann
@@ -18,6 +18,7 @@
#include <vsip/map.hpp>
#include <vsip/impl/setup-assign.hpp>
#include <vsip/impl/working-view.hpp>
+#include <vsip/impl/par-support.hpp>
#include <vsip/impl/par-util.hpp>
Index: src/vsip/support.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/support.hpp,v
retrieving revision 1.28
diff -u -r1.28 support.hpp
--- src/vsip/support.hpp 22 Mar 2006 20:48:58 -0000 1.28
+++ src/vsip/support.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file support.hpp
@author Jules Bergmann
@@ -220,6 +220,7 @@
typedef unsigned int processor_type;
typedef signed int processor_difference_type;
+index_type const no_index = static_cast<index_type>(-1);
index_type const no_subblock = static_cast<index_type>(-1);
processor_type const no_processor = static_cast<processor_type>(-1);
Index: src/vsip/impl/block-copy.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/block-copy.hpp,v
retrieving revision 1.10
diff -u -r1.10 block-copy.hpp
--- src/vsip/impl/block-copy.hpp 10 Feb 2006 22:24:01 -0000 1.10
+++ src/vsip/impl/block-copy.hpp 24 Mar 2006 18:55:40 -0000
@@ -87,7 +87,16 @@
static void exec(BlockT& block, value_type const& val)
{
typedef typename Distributed_local_block<BlockT>::type local_block_type;
- Block_fill<Dim, local_block_type>::exec(get_local_block(block), val);
+ typedef typename impl::View_block_storage<local_block_type>::plain_type
+ type;
+
+ if (block.map().subblock() != no_subblock)
+ {
+ // If get_local_block returns a temporary value, we need to copy it.
+ // Other (if it returns a reference), this captures it.
+ type l_block = get_local_block(block);
+ Block_fill<Dim, local_block_type>::exec(l_block, val);
+ }
}
};
Index: src/vsip/impl/block-traits.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/block-traits.hpp,v
retrieving revision 1.18
diff -u -r1.18 block-traits.hpp
--- src/vsip/impl/block-traits.hpp 10 Feb 2006 22:24:01 -0000 1.18
+++ src/vsip/impl/block-traits.hpp 24 Mar 2006 18:55:40 -0000
@@ -224,6 +224,9 @@
/// Check if lhs map is same as rhs block's map.
+/// Two blocks are the same if they distribute a view into subblocks
+/// containing the same elements.
+
template <typename MapT,
typename BlockT>
struct Is_par_same_map
Index: src/vsip/impl/dispatch-assign.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/dispatch-assign.hpp,v
retrieving revision 1.15
diff -u -r1.15 dispatch-assign.hpp
--- src/vsip/impl/dispatch-assign.hpp 10 Feb 2006 22:24:01 -0000 1.15
+++ src/vsip/impl/dispatch-assign.hpp 24 Mar 2006 18:55:40 -0000
@@ -17,6 +17,7 @@
#include <vsip/impl/static_assert.hpp>
#include <vsip/impl/metaprogramming.hpp>
#include <vsip/impl/block-traits.hpp>
+#include <vsip/impl/map-traits.hpp>
#include <vsip/impl/par-expr.hpp>
#include <vsip/impl/expr_serial_dispatch.hpp>
@@ -32,39 +33,6 @@
namespace impl
{
-/// Traits class to determine if a map is serial or not.
-template <typename Map>
-struct Is_local_map
-{ static bool const value = false; };
-
-template <>
-struct Is_local_map<Local_map>
-{ static bool const value = true; };
-
-template <dimension_type Dim>
-struct Is_local_map<Local_or_global_map<Dim> >
-{ static bool const value = true; };
-
-template <typename Map>
-struct Is_global_map
-{ static bool const value = false; };
-
-template <dimension_type Dim>
-struct Is_global_map<Global_map<Dim> >
-{ static bool const value = true; };
-
-template <dimension_type Dim>
-struct Is_global_map<Local_or_global_map<Dim> >
-{ static bool const value = true; };
-
-template <typename Dist0,
- typename Dist1,
- typename Dist2>
-struct Is_global_map<Map<Dist0, Dist1, Dist2> >
-{ static bool const value = true; };
-
-
-
// Tags used by Dispatch_assign to select assignment implementation.
struct Tag_illegal_mix_of_local_and_global_in_assign;
Index: src/vsip/impl/dist.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/dist.hpp,v
retrieving revision 1.12
diff -u -r1.12 dist.hpp
--- src/vsip/impl/dist.hpp 19 Dec 2005 16:08:55 -0000 1.12
+++ src/vsip/impl/dist.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/impl/dist.hpp
@author Jules Bergmann
@@ -273,6 +273,10 @@
const VSIP_NOTHROW
{ assert(i < dom.size()); return 0; }
+ index_type impl_patch_from_index(Domain<1> const& dom, index_type i)
+ const VSIP_NOTHROW
+ { assert(i < dom.size()); return 0; }
+
index_type impl_local_from_global_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW
{ assert(i < dom.size()); return i; }
@@ -329,6 +333,10 @@
index_type impl_subblock_from_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW;
+ index_type impl_patch_from_index(Domain<1> const& dom, index_type i)
+ const VSIP_NOTHROW
+ { assert(i < dom.size()); return 0; }
+
index_type impl_local_from_global_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW;
@@ -386,6 +394,9 @@
index_type impl_subblock_from_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW;
+ index_type impl_patch_from_index(Domain<1> const& dom, index_type i)
+ const VSIP_NOTHROW;
+
index_type impl_local_from_global_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW;
@@ -484,7 +495,7 @@
// cleanly, then the remaining elements (called the "spill_over")
// are distributed 1 each to first subblocks.
-inline index_type
+inline index_type
Block_dist::impl_subblock_from_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW
{
@@ -499,7 +510,7 @@
-inline index_type
+inline index_type
Block_dist::impl_local_from_global_index(Domain<1> const& dom, index_type i)
const VSIP_NOTHROW
{
@@ -522,7 +533,7 @@
/// I is an index into the subblock.
/// Returns:
-inline index_type
+inline index_type
Block_dist::impl_global_from_local_index(
Domain<1> const& dom,
index_type sb,
@@ -661,20 +672,56 @@
-inline index_type
-Cyclic_dist::impl_subblock_from_index(Domain<1> const& /*dom*/, index_type /*i*/)
+inline index_type
+Cyclic_dist::impl_subblock_from_index(
+ Domain<1> const& /*dom*/,
+ index_type i)
const VSIP_NOTHROW
{
- VSIP_IMPL_THROW(impl::unimplemented("Cyclic_dist::impl_subblock_from_index()"));
+ // Determine global_patch containing index i.
+ index_type p_g = i / contiguity_;
+
+ // Determine subblock holding this patch.
+ index_type sb = p_g % num_subblocks_;
+
+ return sb;
}
-inline index_type
-Cyclic_dist::impl_local_from_global_index(Domain<1> const& /*dom*/, index_type /*i*/)
+inline index_type
+Cyclic_dist::impl_patch_from_index(
+ Domain<1> const& /*dom*/,
+ index_type i)
const VSIP_NOTHROW
{
- VSIP_IMPL_THROW(impl::unimplemented("Cyclic_dist::impl_local_from_global_index()"));
+ // Determine global_patch containing index i.
+ index_type p_g = i / contiguity_;
+
+ // Determine subblock holding this patch.
+ index_type sb = p_g % num_subblocks_;
+ index_type p_l = (p_g - sb) / num_subblocks_;
+
+ return p_l;
+}
+
+
+
+inline index_type
+Cyclic_dist::impl_local_from_global_index(
+ Domain<1> const& /*dom*/,
+ index_type i)
+ const VSIP_NOTHROW
+{
+ // Determine global_patch containing index i at offset.
+ index_type p_g = i / contiguity_;
+ index_type p_offset = i % contiguity_;
+
+ // Convert this global patch to a subblock-patch.
+ index_type sb = p_g % num_subblocks_;
+ index_type p_l = (p_g - sb) / num_subblocks_;
+
+ return p_l * contiguity_ + p_offset;
}
@@ -688,7 +735,7 @@
/// Returns:
/// The global index corresponding to subblock SB index I.
-inline index_type
+inline index_type
Cyclic_dist::impl_global_from_local_index(
Domain<1> const& /*dom*/,
index_type sb,
Index: src/vsip/impl/distributed-block.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/distributed-block.hpp,v
retrieving revision 1.19
diff -u -r1.19 distributed-block.hpp
--- src/vsip/impl/distributed-block.hpp 22 Mar 2006 20:48:58 -0000 1.19
+++ src/vsip/impl/distributed-block.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/impl/distributed-block.hpp
@author Jules Bergmann
@@ -61,16 +61,16 @@
public:
Distributed_block(Domain<dim> const& dom, Map const& map = Map())
: map_ (map),
- proc_ (map_.impl_rank()),
+ proc_ (local_processor()),
sb_ (map_.subblock(proc_)),
subblock_ (NULL)
{
- map_.apply(dom);
+ map_.impl_apply(dom);
for (dimension_type d=0; d<dim; ++d)
size_[d] = dom[d].length();
Domain<dim> sb_dom =
- (sb_ != no_subblock) ? map_.template subblock_domain<dim>(sb_)
+ (sb_ != no_subblock) ? map_.template impl_subblock_domain<dim>(sb_)
: empty_domain<dim>();
subblock_ = new Block(sb_dom);
@@ -79,16 +79,16 @@
Distributed_block(Domain<dim> const& dom, value_type value,
Map const& map = Map())
: map_ (map),
- proc_ (map_.impl_rank()),
+ proc_ (local_processor()),
sb_ (map_.subblock(proc_)),
subblock_ (NULL)
{
- map_.apply(dom);
+ map_.impl_apply(dom);
for (dimension_type d=0; d<dim; ++d)
size_[d] = dom[d].length();
Domain<dim> sb_dom =
- (sb_ != no_subblock) ? map_.template subblock_domain<dim>(sb_)
+ (sb_ != no_subblock) ? map_.template impl_subblock_domain<dim>(sb_)
: empty_domain<dim>();
subblock_ = new Block(sb_dom, value);
@@ -99,16 +99,16 @@
value_type* const ptr,
Map const& map = Map())
: map_ (map),
- proc_ (map_.impl_rank()),
+ proc_ (local_processor()),
sb_ (map_.subblock()),
subblock_ (NULL)
{
- map_.apply(dom);
+ map_.impl_apply(dom);
for (dimension_type d=0; d<dim; ++d)
size_[d] = dom[d].length();
Domain<dim> sb_dom =
- (sb_ != no_subblock) ? map_.template subblock_domain<dim>(sb_)
+ (sb_ != no_subblock) ? map_.template impl_subblock_domain<dim>(sb_)
: empty_domain<dim>();
subblock_ = new Block(sb_dom, ptr);
Index: src/vsip/impl/global_map.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/global_map.hpp,v
retrieving revision 1.9
diff -u -r1.9 global_map.hpp
--- src/vsip/impl/global_map.hpp 11 Jan 2006 16:22:44 -0000 1.9
+++ src/vsip/impl/global_map.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/global_map.hpp
@author Jules Bergmann
@@ -17,7 +17,7 @@
#include <vsip/impl/vector-iterator.hpp>
#include <vsip/impl/par-services.hpp>
#include <vsip/impl/map-traits.hpp>
-#include <vsip/impl/par-util.hpp>
+#include <vsip/impl/par-support.hpp>
#include <vsip/map_fwd.hpp>
@@ -43,7 +43,9 @@
// Accessors.
public:
- length_type num_subblocks() const VSIP_NOTHROW { return 1; }
+ length_type num_subblocks() const VSIP_NOTHROW { return 1; }
+ length_type num_processors() const VSIP_NOTHROW
+ { return vsip::num_processors(); }
index_type subblock(processor_type /*pr*/) const VSIP_NOTHROW
{ return 0; }
@@ -67,26 +69,31 @@
// Applied map functions.
public:
- length_type num_patches (index_type sb) const VSIP_NOTHROW
+ length_type impl_num_patches(index_type sb) const VSIP_NOTHROW
{ assert(sb == 0); return 1; }
- void apply(Domain<Dim> const& dom) VSIP_NOTHROW
+ void impl_apply(Domain<Dim> const& dom) VSIP_NOTHROW
{ dom_ = dom; }
template <dimension_type Dim2>
- Domain<Dim2> subblock_domain(index_type sb) const VSIP_NOTHROW
+ Domain<Dim2> impl_subblock_domain(index_type sb) const VSIP_NOTHROW
{ assert(sb == 0); return dom_; }
template <dimension_type Dim2>
- Domain<Dim2> global_domain(index_type sb, index_type patch)
+ Domain<Dim2> impl_global_domain(index_type sb, index_type patch)
const VSIP_NOTHROW
{ assert(sb == 0 && patch == 0); return dom_; }
template <dimension_type Dim2>
- Domain<Dim2> local_domain (index_type sb, index_type patch)
+ Domain<Dim2> impl_local_domain (index_type sb, index_type patch)
const VSIP_NOTHROW
{ assert(sb == 0 && patch == 0); return dom_; }
+ index_type impl_global_from_local_index(dimension_type /*d*/, index_type sb,
+ index_type idx)
+ const VSIP_NOTHROW
+ { assert(sb == 0); return idx; }
+
index_type impl_local_from_global_index(dimension_type /*d*/, index_type idx)
const VSIP_NOTHROW
{ return idx; }
@@ -101,19 +108,13 @@
impl::Communicator impl_comm() const { return impl::default_communicator(); }
pvec_type impl_pvec() const
{ return impl::default_communicator().pvec(); }
- processor_type impl_rank() const
- { return impl::default_communicator().rank(); }
- processor_type impl_size() const
- { return impl::default_communicator().size(); }
length_type impl_working_size() const
- { return this->impl_size(); }
+ { return this->num_processors(); }
- processor_type impl_proc(index_type idx) const
+ processor_type impl_proc_from_rank(index_type idx) const
{ return this->impl_pvec()[idx]; }
- length_type impl_num_procs() const { return this->impl_size(); }
-
// Member data.
private:
Domain<Dim> dom_;
@@ -139,8 +140,16 @@
/// Specialize global traits for Local_or_global_map.
template <dimension_type Dim>
-struct Is_global_only<Local_or_global_map<Dim> >
-{ static bool const value = false; };
+struct Is_global_map<Global_map<Dim> >
+{ static bool const value = true; };
+
+template <dimension_type Dim>
+struct Is_local_map<Local_or_global_map<Dim> >
+{ static bool const value = true; };
+
+template <dimension_type Dim>
+struct Is_global_map<Local_or_global_map<Dim> >
+{ static bool const value = true; };
Index: src/vsip/impl/layout.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/layout.hpp,v
retrieving revision 1.18
diff -u -r1.18 layout.hpp
--- src/vsip/impl/layout.hpp 7 Mar 2006 02:15:22 -0000 1.18
+++ src/vsip/impl/layout.hpp 24 Mar 2006 18:55:40 -0000
@@ -185,7 +185,10 @@
index_type index(index_type idx0)
const VSIP_NOTHROW
- { return idx0; }
+ {
+ assert(idx0 < size_[0]);
+ return idx0;
+ }
index_type index(Index<1> idx)
const VSIP_NOTHROW
@@ -242,7 +245,10 @@
index_type index(index_type idx0)
const VSIP_NOTHROW
- { return idx0; }
+ {
+ assert(idx0 < size_[0]);
+ return idx0;
+ }
index_type index(Index<1> idx)
const VSIP_NOTHROW
@@ -293,7 +299,10 @@
index_type index(index_type idx0)
const VSIP_NOTHROW
- { return idx0; }
+ {
+ assert(idx0 < size_[0]);
+ return idx0;
+ }
index_type index(Index<1> idx)
const VSIP_NOTHROW
@@ -351,7 +360,10 @@
index_type index(index_type idx0, index_type idx1)
const VSIP_NOTHROW
- { return idx0 * size_[1] + idx1; }
+ {
+ assert(idx0 < size_[0] && idx1 < size_[1]);
+ return idx0 * size_[1] + idx1;
+ }
index_type index(Index<2> idx)
const VSIP_NOTHROW
@@ -409,7 +421,10 @@
index_type index(index_type idx0, index_type idx1)
const VSIP_NOTHROW
- { return idx0 + idx1 * size_[0]; }
+ {
+ assert(idx0 < size_[0] && idx1 < size_[1]);
+ return idx0 + idx1 * size_[0];
+ }
index_type index(Index<2> idx)
const VSIP_NOTHROW
@@ -484,7 +499,10 @@
index_type index(index_type idx0, index_type idx1)
const VSIP_NOTHROW
- { return idx0 * stride_ + idx1; }
+ {
+ assert(idx0 < size_[0] && idx1 < size_[1]);
+ return idx0 * stride_ + idx1;
+ }
index_type index(Index<2> idx)
const VSIP_NOTHROW
@@ -560,7 +578,10 @@
index_type index(index_type idx0, index_type idx1)
const VSIP_NOTHROW
- { return idx0 + idx1 * stride_; }
+ {
+ assert(idx0 < size_[0] && idx1 < size_[1]);
+ return idx0 + idx1 * stride_;
+ }
index_type index(Index<2> idx)
const VSIP_NOTHROW
@@ -620,6 +641,7 @@
index_type index(Point<3> idx)
const VSIP_NOTHROW
{
+ assert(idx[0] < size_[0] && idx[1] < size_[1] && idx[2] < size_[2]);
return idx[Dim0]*size_[Dim1]*size_[Dim2] +
idx[Dim1]*size_[Dim2] +
idx[Dim2];
@@ -701,6 +723,7 @@
index_type index(Point<3> idx)
const VSIP_NOTHROW
{
+ assert(idx[0] < size_[0] && idx[1] < size_[1] && idx[2] < size_[2]);
return idx[Dim0]*stride_[Dim0] +
idx[Dim1]*stride_[Dim1] +
idx[Dim2];
Index: src/vsip/impl/local_map.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/local_map.hpp,v
retrieving revision 1.6
diff -u -r1.6 local_map.hpp
--- src/vsip/impl/local_map.hpp 5 Dec 2005 19:19:18 -0000 1.6
+++ src/vsip/impl/local_map.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/local_map.hpp
@author Jules Bergmann
@@ -71,11 +71,11 @@
// Applied map functions.
public:
- length_type num_patches (index_type sb) const VSIP_NOTHROW
+ length_type impl_num_patches(index_type sb) const VSIP_NOTHROW
{ assert(sb == 0); return 1; }
template <dimension_type Dim>
- void apply(Domain<Dim> const& /*dom*/) VSIP_NOTHROW
+ void impl_apply(Domain<Dim> const& /*dom*/) VSIP_NOTHROW
{}
template <dimension_type Dim2>
@@ -97,18 +97,12 @@
impl::Communicator impl_comm() const { return impl::default_communicator(); }
pvec_type impl_pvec() const
{ return impl::default_communicator().pvec(); }
- processor_type impl_rank() const
- { return impl::default_communicator().rank(); }
- processor_type impl_size() const
- { return impl::default_communicator().size(); }
length_type impl_working_size() const
- { return this->impl_size(); }
-
- processor_type impl_proc(index_type idx) const
- { return this->impl_pvec()[idx]; }
+ { return 1; }
- length_type impl_num_procs() const { return this->impl_size(); }
+ processor_type impl_proc_from_rank(index_type idx) const
+ { assert(idx == 0); return local_processor(); }
// No member data.
};
@@ -119,13 +113,9 @@
/// Specialize local/global traits for Local_map.
template <>
-struct Is_local_only<Local_map>
+struct Is_local_map<Local_map>
{ static bool const value = true; };
-template <>
-struct Is_global_only<Local_map>
-{ static bool const value = false; };
-
template < >
Index: src/vsip/impl/map-traits.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/map-traits.hpp,v
retrieving revision 1.1
diff -u -r1.1 map-traits.hpp
--- src/vsip/impl/map-traits.hpp 2 Nov 2005 18:44:04 -0000 1.1
+++ src/vsip/impl/map-traits.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/impl/block-traits.hpp
@author Jules Bergmann
@@ -29,12 +29,22 @@
/// Traits class to determine if a map is serial or not.
template <typename MapT>
-struct Is_local_only
+struct Is_local_map
{ static bool const value = false; };
template <typename Map>
+struct Is_global_map
+{ static bool const value = false; };
+
+template <typename MapT>
+struct Is_local_only
+{ static bool const value = Is_local_map<MapT>::value &&
+ !Is_global_map<MapT>::value; };
+
+template <typename MapT>
struct Is_global_only
-{ static bool const value = true; };
+{ static bool const value = Is_global_map<MapT>::value &&
+ !Is_local_map<MapT>::value; };
} // namespace vsip::impl
Index: src/vsip/impl/par-assign.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/par-assign.hpp,v
retrieving revision 1.10
diff -u -r1.10 par-assign.hpp
--- src/vsip/impl/par-assign.hpp 5 Dec 2005 19:19:18 -0000 1.10
+++ src/vsip/impl/par-assign.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/impl/par-assign.hpp
@author Jules Bergmann
@@ -76,9 +76,6 @@
dst_appmap_t const& dst_am = dst_.block().map();
assert(src_am.impl_comm() == dst_am.impl_comm());
- // Implies:
- assert(src_am.impl_rank() == dst_am.impl_rank());
- assert(src_am.impl_size() == dst_am.impl_size());
}
void operator()()
@@ -89,7 +86,7 @@
dst_appmap_t const& dst_am = dst_.block().map();
- processor_type rank = src_am.impl_rank();
+ processor_type rank = local_processor();
src_sb_iterator cur = src_am.subblocks_begin(rank);
src_sb_iterator end = src_am.subblocks_end(rank);
@@ -249,9 +246,6 @@
dst_appmap_t const& dst_am = dst_.block().map();
assert(src_am.impl_comm() == dst_am.impl_comm());
- // Implies:
- assert(src_am.impl_rank() == dst_am.impl_rank());
- assert(src_am.impl_size() == dst_am.impl_size());
}
void operator()()
@@ -261,9 +255,7 @@
src_appmap_t const& src_am = src_.block().map();
dst_appmap_t const& dst_am = dst_.block().map();
- assert(src_am.impl_rank() == dst_am.impl_rank());
-
- processor_type rank = src_am.impl_rank();
+ processor_type rank = local_processor();
impl::Communicator comm = src_am.impl_comm();
@@ -485,9 +477,9 @@
: dst_ (dst),
src_ (src.block()),
send_list (),
- send_size (dst_.block().map().impl_size()), // reserve comm size
+ send_size (num_processors()), // reserve comm size
recv_list (),
- recv_size (src_.block().map().impl_size()), // reserve comm size
+ recv_size (num_processors()), // reserve comm size
sbuf_list (),
req_list (),
msg_count (0)
@@ -496,11 +488,6 @@
dst_appmap_t const& dst_am = dst_.block().map();
assert(src_am.impl_comm() == dst_am.impl_comm());
- // Implies:
- assert(src_am.impl_rank() == dst_am.impl_rank());
- assert(src_am.impl_size() == dst_am.impl_size());
-
- assert(src_am.impl_rank() == dst_am.impl_rank());
build_send_list();
build_recv_list();
@@ -518,9 +505,7 @@
src_appmap_t const& src_am = src_.block().map();
dst_appmap_t const& dst_am = dst_.block().map();
- assert(src_am.impl_rank() == dst_am.impl_rank());
-
- processor_type rank = src_am.impl_rank();
+ processor_type rank = local_processor();
// Iterate over all processors
for (index_type pi=0; pi<dst_am.impl_pvec().size(); ++pi)
@@ -595,9 +580,7 @@
src_appmap_t const& src_am = src_.block().map();
dst_appmap_t const& dst_am = dst_.block().map();
- assert(src_am.impl_rank() == dst_am.impl_rank());
-
- processor_type rank = src_am.impl_rank();
+ processor_type rank = local_processor();
for (index_type pi=0; pi<src_am.impl_pvec().size(); ++pi)
recv_size[pi] = 0;
@@ -668,8 +651,8 @@
dst_appmap_t dst_am)
{
impl::Communicator comm = dst_am.impl_comm();
- processor_type rank = dst_am.impl_rank();
- length_type size = dst_am.impl_size();
+ processor_type rank = local_processor();
+ length_type size = num_processors();
for (index_type pi=0; pi<dst_am.impl_pvec().size(); ++pi)
{
@@ -721,8 +704,8 @@
src_appmap_t src_am)
{
impl::Communicator comm = src_am.impl_comm();
- processor_type rank = src_am.impl_rank();
- length_type size = src_am.impl_size();
+ processor_type rank = local_processor();
+ length_type size = num_processors();
// Iterate over all sending processors
for (index_type pi=0; pi<src_am.impl_pvec().size(); ++pi)
Index: src/vsip/impl/par-chain-assign.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/par-chain-assign.hpp,v
retrieving revision 1.17
diff -u -r1.17 par-chain-assign.hpp
--- src/vsip/impl/par-chain-assign.hpp 3 Mar 2006 14:30:53 -0000 1.17
+++ src/vsip/impl/par-chain-assign.hpp 24 Mar 2006 18:55:40 -0000
@@ -128,7 +128,7 @@
typedef typename View_of_dim<Dim, T, local_block_type>::const_type
local_view_type;
- processor_type rank = am.impl_rank();
+ processor_type rank = local_processor();
// First set all subblock ext pointers to NULL.
length_type tot_sb = am.num_subblocks();
@@ -280,11 +280,6 @@
{
impl::profile::Scope_event ev("Chained_parallel_assign-cons");
assert(src_am_.impl_comm() == dst_am_.impl_comm());
- // Implies:
- assert(src_am_.impl_rank() == dst_am_.impl_rank());
- assert(src_am_.impl_size() == dst_am_.impl_size());
-
- assert(src_am_.impl_rank() == dst_am_.impl_rank());
par_chain_assign::build_ext_array<Dim, T2, Block2>(
src_, src_am_, src_ext_, impl::SYNC_IN);
@@ -420,25 +415,27 @@
Chained_parallel_assign<Dim, T1, T2, Block1, Block2>::build_send_list()
{
impl::profile::Scope_event ev("Chained_parallel_assign-build_send_list");
- processor_type rank = src_am_.impl_rank();
+ processor_type rank = local_processor();
#if VSIPL_IMPL_PCA_ROTATE
- index_type offset = src_am_.lookup_index(rank);
+ index_type offset = src_am_.impl_rank_from_proc(rank);
#endif
length_type dsize = dst_am_.impl_working_size();
// std::min(dst_am_.num_subblocks(), dst_am_.impl_pvec().size());
-#if VSIP_IMPL_PCA_VERBOSE
- std::cout << "\n"
- << "(" << rank << ") "
- << "build_send_list() -------------------------------------\n"
- << "(" << rank << ") "
- << " dsize = " << dsize << std::endl;
+#if VSIP_IMPL_PCA_VERBOSE >= 1
+ std::cout << "(" << rank << ") "
+ << "build_send_list(dsize: " << dsize
+ << ") -------------------------------------\n";
#endif
index_type src_sb = src_am_.subblock(rank);
- if (src_sb != no_subblock)
+ // If multiple processors have the subblock, the first processor
+ // is responsible for sending it.
+
+ if (src_sb != no_subblock &&
+ *(src_am_.processor_begin(src_sb)) == rank)
{
// Iterate over all processors
for (index_type pi=0; pi<dsize; ++pi)
@@ -447,9 +444,9 @@
// (Currently does not work, it needs to take into account the
// number of subblocks).
#if VSIPL_IMPL_PCA_ROTATE
- processor_type proc = dst_am_.impl_proc((pi + offset) % dsize);
+ processor_type proc = dst_am_.impl_proc_from_rank((pi + offset) % dsize);
#else
- processor_type proc = dst_am_.impl_proc(pi);
+ processor_type proc = dst_am_.impl_proc_from_rank(pi);
#endif
// Transfers that stay on this processor is handled by the copy_list.
@@ -468,14 +465,14 @@
src_ext_type* ext = src_ext_[src_sb];
ext->begin();
- for (index_type dp=0; dp<dst_am_.num_patches(dst_sb); ++dp)
+ for (index_type dp=0; dp<num_patches(dst_, dst_sb); ++dp)
{
- Domain<dim> dst_dom = dst_am_.template global_domain<dim>(dst_sb, dp);
+ Domain<dim> dst_dom = global_domain(dst_, dst_sb, dp);
- for (index_type sp=0; sp<src_am_.num_patches(src_sb); ++sp)
+ for (index_type sp=0; sp<num_patches(src_, src_sb); ++sp)
{
- Domain<dim> src_dom = src_am_.template global_domain<dim>(src_sb, sp);
- Domain<dim> src_ldom = src_am_.template local_domain<dim>(src_sb, sp);
+ Domain<dim> src_dom = global_domain(src_, src_sb, sp);
+ Domain<dim> src_ldom = local_domain(src_, src_sb, sp);
Domain<dim> intr;
@@ -487,7 +484,7 @@
par_chain_assign::chain_add<dst_order_t>(builder, *ext, send_dom);
-#if VSIP_IMPL_PCA_VERBOSE
+#if VSIP_IMPL_PCA_VERBOSE >= 2
std::cout << "(" << rank << ") send "
<< rank << "/" << src_sb << "/" << sp
<< " -> "
@@ -524,18 +521,17 @@
Chained_parallel_assign<Dim, T1, T2, Block1, Block2>::build_recv_list()
{
impl::profile::Scope_event ev("Chained_parallel_assign-build_recv_list");
- processor_type rank = src_am_.impl_rank();
+ processor_type rank = local_processor();
#if VSIPL_IMPL_PCA_ROTATE
- index_type offset = dst_am_.lookup_index(rank);
+ index_type offset = dst_am_.impl_rank_from_proc(rank);
#endif
length_type ssize = src_am_.impl_working_size();
-#if VSIP_IMPL_PCA_VERBOSE
- std::cout << "\n"
- << "(" << rank << ")"
- << "build_recv_list() -------------------------------------\n"
- << " ssize = " << ssize << std::endl;
+#if VSIP_IMPL_PCA_VERBOSE >= 1
+ std::cout << "(" << rank << ") "
+ << "build_recv_list(ssize: " << ssize
+ << ") -------------------------------------\n";
#endif
index_type dst_sb = dst_am_.subblock(rank);
@@ -551,8 +547,8 @@
// Rotate message order so processors don't all send to 0,
// then 1, etc (Currently does not work, it needs to take into
// account the number of subblocks).
- // processor_type proc = (src_am_.impl_proc(pi) + rank) % size;
- processor_type proc = src_am_.impl_proc(pi);
+ // processor_type proc = (src_am_.impl_proc_from_rank(pi) + rank) % size;
+ processor_type proc = src_am_.impl_proc_from_rank(pi);
// Transfers that stay on this processor is handled by the copy_list.
if (!disable_copy && proc == rank)
@@ -562,20 +558,24 @@
index_type src_sb = src_am_.subblock(proc);
- if (src_sb != no_subblock)
+ // If multiple processors have the subblock, the first processor
+ // is responsible for sending it to us.
+
+ if (src_sb != no_subblock &&
+ *(src_am_.processor_begin(src_sb)) == proc)
{
// Check to see if destination processor already has block
if (!disable_copy && processor_has_block(src_am_, rank, src_sb))
continue;
- for (index_type dp=0; dp<dst_am_.num_patches(dst_sb); ++dp)
+ for (index_type dp=0; dp<num_patches(dst_, dst_sb); ++dp)
{
- Domain<dim> dst_dom = dst_am_.template global_domain<dim>(dst_sb, dp);
- Domain<dim> dst_ldom = dst_am_.template local_domain<dim>(dst_sb, dp);
+ Domain<dim> dst_dom = global_domain(dst_, dst_sb, dp);
+ Domain<dim> dst_ldom = local_domain(dst_, dst_sb, dp);
- for (index_type sp=0; sp<src_am_.num_patches(src_sb); ++sp)
+ for (index_type sp=0; sp<num_patches(src_, src_sb); ++sp)
{
- Domain<dim> src_dom = src_am_.template global_domain<dim>(src_sb, sp);
+ Domain<dim> src_dom = global_domain(src_, src_sb, sp);
Domain<dim> intr;
@@ -587,7 +587,7 @@
par_chain_assign::chain_add<dst_order_t>(builder, *ext, recv_dom);
-#if VSIP_IMPL_PCA_VERBOSE
+#if VSIP_IMPL_PCA_VERBOSE >= 1
std::cout << "(" << rank << ") recv "
<< rank << "/" << dst_sb << "/" << dp
<< " <- "
@@ -621,13 +621,12 @@
Chained_parallel_assign<Dim, T1, T2, Block1, Block2>::build_copy_list()
{
impl::profile::Scope_event ev("Chained_parallel_assign-build_copy_list");
- processor_type rank = src_am_.impl_rank();
+ processor_type rank = local_processor();
-#if VSIP_IMPL_PCA_VERBOSE
- std::cout << "\n"
- << "(" << rank << ")"
- << "build_copy_list() -------------------------------------\n"
- << " num_procs = " << src_am_.impl_num_procs() << std::endl;
+#if VSIP_IMPL_PCA_VERBOSE >= 1
+ std::cout << "(" << rank << ") "
+ << "build_copy_list(num_procs: " << src_am_.num_processors()
+ << ") -------------------------------------\n";
#endif
index_type dst_sb = dst_am_.subblock(rank);
@@ -637,19 +636,19 @@
index_type src_sb = src_am_.subblock(rank);
if (src_sb != no_subblock)
{
- for (index_type dp=0; dp<dst_am_.num_patches(dst_sb); ++dp)
+ for (index_type dp=0; dp<num_patches(dst_, dst_sb); ++dp)
{
- Domain<dim> dst_dom = dst_am_.template global_domain<dim>(dst_sb, dp);
- Domain<dim> dst_ldom = dst_am_.template local_domain<dim> (dst_sb, dp);
+ Domain<dim> dst_dom = global_domain(dst_, dst_sb, dp);
+ Domain<dim> dst_ldom = local_domain (dst_, dst_sb, dp);
- for (index_type sp=0; sp<src_am_.num_patches(src_sb); ++sp)
+ for (index_type sp=0; sp<num_patches(src_, src_sb); ++sp)
{
- Domain<dim> src_dom = src_am_.template global_domain<dim>(src_sb, sp);
- Domain<dim> src_ldom = src_am_.template local_domain<dim> (src_sb, sp);
+ Domain<dim> src_dom = global_domain(src_, src_sb, sp);
+ Domain<dim> src_ldom = local_domain (src_, src_sb, sp);
Domain<dim> intr;
-#if VSIP_IMPL_PCA_VERBOSE
+#if VSIP_IMPL_PCA_VERBOSE >= 2
// std::cout << " - dst " << dst_sb << "/" << dp << std::endl
// << " src " << src_sb << "/" << sp << std::endl
// ;
@@ -666,8 +665,9 @@
copy_list.push_back(Copy_record(src_sb, dst_sb, send_dom, recv_dom));
-#if VSIP_IMPL_PCA_VERBOSE
- std::cout << "copy src: " << src_sb << "/" << sp
+#if VSIP_IMPL_PCA_VERBOSE >= 2
+ std::cout << "(" << rank << ")"
+ << "copy src: " << src_sb << "/" << sp
<< " " << send_dom
<< " dst: " << dst_sb << "/" << dp
<< " " << recv_dom
@@ -695,12 +695,11 @@
impl::profile::Scope_event ev("Chained_parallel_assign-exec_send_list");
impl::Communicator comm = dst_am_.impl_comm();
-#if VSIP_IMPL_PCA_VERBOSE
- processor_type rank = dst_am_.impl_rank();
- std::cout << "\n"
- << "(" << rank << ")"
- << "exec_send_list() -------------------------------------\n"
- << " size = " << send_list.size() << std::endl;
+#if VSIP_IMPL_PCA_VERBOSE >= 1
+ processor_type rank = local_processor();
+ std::cout << "(" << rank << ") "
+ << "exec_send_list(size: " << send_list.size()
+ << ") -------------------------------------\n";
#endif
length_type dsize = dst_am_.impl_working_size();
@@ -709,7 +708,7 @@
for (index_type pi=0; pi<dsize; ++pi)
{
// TODO: Rotate
- processor_type proc = dst_am_.impl_proc(pi);
+ processor_type proc = dst_am_.impl_proc_from_rank(pi);
impl::Chain_builder builder;
@@ -754,12 +753,11 @@
impl::profile::Scope_event ev("Chained_parallel_assign-exec_recv_list");
impl::Communicator comm = src_am_.impl_comm();
-#if VSIP_IMPL_PCA_VERBOSE
- processor_type rank = src_am_.impl_rank();
- std::cout << "\n"
- << "(" << rank << ")"
- << "exec_recv_list() -------------------------------------\n"
- << " size = " << recv_list.size() << std::endl;
+#if VSIP_IMPL_PCA_VERBOSE >= 1
+ processor_type rank = local_processor();
+ std::cout << "(" << rank << ") "
+ << "exec_recv_list(size: " << recv_list.size()
+ << ") -------------------------------------\n";
#endif
length_type ssize = src_am_.impl_working_size();
@@ -767,8 +765,8 @@
// Iterate over all sending processors
for (index_type pi=0; pi<ssize; ++pi)
{
- // processor_type proc = (src_am_.impl_proc(pi) + rank) % size;
- processor_type proc = src_am_.impl_proc(pi);
+ // processor_type proc = (src_am_.impl_proc_from_rank(pi) + rank) % size;
+ processor_type proc = src_am_.impl_proc_from_rank(pi);
impl::Chain_builder builder;
@@ -813,12 +811,11 @@
impl::profile::Scope_event ev("Chained_parallel_assign-exec_copy_list");
impl::Communicator comm = dst_am_.impl_comm();
-#if VSIP_IMPL_PCA_VERBOSE
- processor_type rank = dst_am_.impl_rank();
- std::cout << "\n"
- << "(" << rank << ")"
- << "exec_copy_list() -------------------------------------\n"
- << " size = " << copy_list.size() << std::endl;
+#if VSIP_IMPL_PCA_VERBOSE >= 1
+ processor_type rank = local_processor();
+ std::cout << "(" << rank << ") "
+ << "exec_copy_list(size: " << copy_list.size()
+ << ") -------------------------------------\n";
#endif
typedef typename std::vector<Copy_record>::iterator cl_iterator;
@@ -832,8 +829,9 @@
dst_lview_type dst_lview = get_local_view(dst_);
dst_lview((*cl_cur).dst_dom_) = src_lview((*cl_cur).src_dom_);
-#if VSIP_IMPL_PCA_VERBOSE
- std::cout << "src subblock: " << (*cl_cur).src_sb_ << " -> "
+#if VSIP_IMPL_PCA_VERBOSE >= 2
+ std::cout << "(" << rank << ") "
+ << "src subblock: " << (*cl_cur).src_sb_ << " -> "
<< "dst subblock: " << (*cl_cur).dst_sb_ << std::endl
<< dst_lview((*cl_cur).dst_dom_);
#endif
Index: src/vsip/impl/par-support.hpp
===================================================================
RCS file: src/vsip/impl/par-support.hpp
diff -N src/vsip/impl/par-support.hpp
--- /dev/null 1 Jan 1970 00:00:00 -0000
+++ src/vsip/impl/par-support.hpp 24 Mar 2006 18:55:40 -0000
@@ -0,0 +1,515 @@
+/* Copyright (c) 2006 by CodeSourcery, LLC. All rights reserved. */
+
+/** @file vsip/impl/par-support.hpp
+ @author Jules Bergmann
+ @date 2006-03-14
+ @brief VSIPL++ Library: Parallel support funcions [view.support.fcn].
+
+*/
+
+#ifndef VSIP_IMPL_PAR_SUPPORT_HPP
+#define VSIP_IMPL_PAR_SUPPORT_HPP
+
+/***********************************************************************
+ Included Files
+***********************************************************************/
+
+#include <vsip/impl/par-services.hpp>
+#include <vsip/support.hpp>
+#include <vsip/vector.hpp>
+#include <vsip/matrix.hpp>
+#include <vsip/impl/distributed-block.hpp>
+#include <vsip/impl/domain-utils.hpp>
+
+
+
+/***********************************************************************
+ Definitions
+***********************************************************************/
+
+namespace vsip
+{
+
+const_Vector<processor_type> processor_set();
+
+namespace impl
+{
+
+namespace psf_detail
+{
+
+// Return the subdomain of a view/map pair for a subblock.
+
+template <typename ViewT>
+inline Domain<ViewT::dim>
+subblock_domain(
+ ViewT const& view,
+ Local_map const& /*map*/,
+ index_type sb)
+{
+ assert(sb == 0 || sb == no_subblock);
+ return (sb == 0) ? block_domain<ViewT::dim>(view.block())
+ : empty_domain<ViewT::dim>();
+}
+
+template <typename ViewT,
+ typename MapT>
+inline Domain<ViewT::dim>
+subblock_domain(
+ ViewT const& /*view*/,
+ MapT const& map,
+ index_type sb)
+{
+ return map.template impl_subblock_domain<ViewT::dim>(sb);
+}
+
+
+
+// Return the local domain of a view/map pair for a subblock/patch.
+
+template <typename ViewT>
+inline Domain<ViewT::dim>
+local_domain(
+ ViewT const& view,
+ Local_map const& /*map*/,
+ index_type sb,
+ index_type p)
+{
+ assert((sb == 0 && p == 0) || sb == no_subblock);
+ return (sb == 0) ? block_domain<ViewT::dim>(view.block())
+ : empty_domain<ViewT::dim>();
+}
+
+template <typename ViewT,
+ typename MapT>
+inline Domain<ViewT::dim>
+local_domain(
+ ViewT const& /*view*/,
+ MapT const& map,
+ index_type sb,
+ index_type p)
+{
+ return map.template impl_local_domain<ViewT::dim>(sb, p);
+}
+
+
+
+// Return the global domain of a view/map pair for a subblock/patch.
+
+template <typename ViewT>
+inline Domain<ViewT::dim>
+global_domain(
+ ViewT const& view,
+ Local_map const& /*map*/,
+ index_type sb,
+ index_type p)
+{
+ assert((sb == 0 && p == 0) || sb == no_subblock);
+ return (sb == 0) ? block_domain<ViewT::dim>(view.block())
+ : empty_domain<ViewT::dim>();
+}
+
+template <typename ViewT,
+ typename MapT>
+inline Domain<ViewT::dim>
+global_domain(
+ ViewT const& /*view*/,
+ MapT const& map,
+ index_type sb,
+ index_type p)
+{
+ return map.template impl_global_domain<ViewT::dim>(sb, p);
+}
+
+} // namespace vsip::impl::psf_detail
+} // namespace vsip::impl
+
+
+
+/***********************************************************************
+ Definitions - [view.support.fcn] parallel support functions
+***********************************************************************/
+
+/// Return the domain of VIEW's subblock SB.
+
+/// Requires
+/// VIEW to be a view
+/// SB to either be a valid subblock of VIEW, or the value no_subblock.
+///
+/// Returns
+/// The domain of VIEW's subblock SB if SB is valid, the empty
+/// domain if SB == no_subblock.
+
+template <typename ViewT>
+Domain<ViewT::dim>
+subblock_domain(
+ ViewT const& view,
+ index_type sb)
+{
+ return impl::psf_detail::subblock_domain(view, view.block().map(), sb);
+}
+
+
+
+/// Return the domain of VIEW's subblock held by the local processor.
+
+/// Requires
+/// VIEW to be a view
+///
+/// Returns
+/// The domain of VIEW's subblock held by the local processor.
+
+template <typename ViewT>
+Domain<ViewT::dim>
+subblock_domain(
+ ViewT const& view)
+{
+ return impl::psf_detail::subblock_domain(view, view.block().map(),
+ view.block().map().subblock());
+}
+
+
+
+/// Return the local domain of VIEW's subblock SB patch P
+
+/// Requires
+/// VIEW to be a view
+/// SB to either be a valid subblock of VIEW, or the value no_subblock.
+/// P to either be a valid patch of subblock SB.
+///
+/// Returns
+/// The local domain of VIEW's subblock SB patch P if SB is valid,
+/// the empty domain if SB == no_subblock.
+
+template <typename ViewT>
+Domain<ViewT::dim>
+local_domain(
+ ViewT const& view,
+ index_type sb,
+ index_type p)
+{
+ return impl::psf_detail::local_domain(view, view.block().map(), sb, p);
+}
+
+
+
+/// Return the local domain of VIEW's patch P on the local processor's subblock
+
+/// Requires
+/// VIEW to be a view
+/// P to either be a valid patch of the local processor's subblock.
+///
+/// Returns
+/// The local domain of VIEW's patch P of the local processor's subblock
+/// if the local processor holds a subblock,
+/// The empty domain otherwise.
+
+template <typename ViewT>
+Domain<ViewT::dim>
+local_domain(
+ ViewT const& view,
+ index_type p=0)
+{
+ return impl::psf_detail::local_domain(view, view.block().map(),
+ view.block().map().subblock(),
+ p);
+}
+
+
+
+/// Return the global domain of VIEW's subblock SB patch P
+
+/// Requires
+/// VIEW to be a view
+/// SB to either be a valid subblock of VIEW, or the value no_subblock.
+/// P to either be a valid patch of subblock SB.
+///
+/// Returns
+/// The global domain of VIEW's subblock SB patch P if SB is valid,
+/// the empty domain if SB == no_subblock.
+
+template <typename ViewT>
+Domain<ViewT::dim>
+global_domain(
+ ViewT const& view,
+ index_type sb,
+ index_type p)
+{
+ return impl::psf_detail::global_domain(view, view.block().map(), sb, p);
+}
+
+
+
+/// Return the global domain of VIEW's local subblock patch P
+
+/// Requires
+/// VIEW to be a view
+/// P to either be a valid patch of the local processor's subblock.
+///
+/// Returns
+/// The global domain of VIEW's patch P of the local processor's subblock
+/// if the local processor holds a subblock,
+/// The empty domain otherwise.
+
+template <typename ViewT>
+Domain<ViewT::dim>
+global_domain(
+ ViewT const& view,
+ index_type p=0)
+{
+ return impl::psf_detail::global_domain(view, view.block().map(),
+ view.block().map().subblock(),
+ p);
+}
+
+
+/// Return the number of subblocks VIEW is distrubted over.
+
+/// Requires
+/// VIEW to be a view
+
+template <typename ViewT>
+length_type
+num_subblocks(
+ ViewT const& view)
+{
+ return view.block().map().num_subblocks();
+}
+
+
+
+/// Return the number of patches in VIEW's subblock SB.
+
+/// Requires
+/// VIEW to be a view.
+/// SB to either be a valid subblock of VIEW, or the value no_subblock.
+
+template <typename ViewT>
+length_type
+num_patches(
+ ViewT const& view,
+ index_type sb)
+{
+ return view.block().map().impl_num_patches(sb);
+}
+
+
+
+/// Return the number of patches in VIEW's subblock held on the local
+/// processor.
+
+/// Requires
+/// VIEW to be a view.
+
+template <typename ViewT>
+length_type
+num_patches(
+ ViewT const& view)
+{
+ return view.block().map().impl_num_patches(view.block().map().subblock());
+}
+
+
+
+/// Return the subblock rank held by processor PR.
+
+/// Requires
+/// VIEW to be a view.
+/// PR to be processor.
+///
+/// Returns
+/// The subblock rank of VIEW held by processor PR if it holds a subblock,
+/// NO_SUBBLOCK otherwise.
+
+template <typename ViewT>
+index_type
+subblock(
+ ViewT const& view,
+ processor_type pr)
+{
+ return view.block().map().subblock(pr);
+}
+
+
+
+/// Return the subblock rank held by processor PR.
+
+/// Requires
+/// VIEW to be a view.
+///
+/// Returns
+/// The subblock rank of VIEW held by local processor,
+/// or NO_SUBBLOCK if it does not hold a subblock.
+
+template <typename ViewT>
+index_type
+subblock(
+ ViewT const& view)
+{
+ return view.block().map().subblock();
+}
+
+
+
+/// Determine which subblock holds VIEW's global index IDX
+
+template <typename ViewT>
+index_type
+subblock_from_global_index(
+ ViewT const& view,
+ Index<ViewT::dim> const& idx)
+{
+ for (dimension_type d=0; d<ViewT::dim; ++d)
+ assert(idx[d] < view.size(d));
+
+ return view.block().
+ map().template impl_subblock_from_global_index<ViewT::dim>(idx);
+}
+
+
+
+/// Determine which patch holds VIEW's global index IDX
+
+/// Notes:
+/// This patch is only valid in the subblock returned by
+/// subblock_from_global_index.
+
+template <typename ViewT>
+index_type
+patch_from_global_index(
+ ViewT const& view,
+ Index<ViewT::dim> const& idx)
+{
+ for (dimension_type d=0; d<ViewT::dim; ++d)
+ assert(idx[d] < view.size(d));
+
+ return view.block().
+ map().template impl_subblock_from_global_index<ViewT::dim>(idx);
+}
+
+
+
+/***********************************************************************
+ local_from_global_index
+***********************************************************************/
+
+/// Determine the local index corresponding to VIEW's global index G_IDX.
+
+/// Notes:
+/// This local index is only valid in processors hold the subblock
+/// returned by subblock_from_global_index.
+
+template <typename ViewT>
+Index<ViewT::dim>
+local_from_global_index(
+ ViewT const& view,
+ Index<ViewT::dim> const& g_idx)
+VSIP_NOTHROW
+{
+ Index<ViewT::dim> l_idx;
+
+ for (dimension_type d=0; d<ViewT::dim; ++d)
+ l_idx[d] = local_from_global_index(view, d, g_idx[d]);
+
+ return l_idx;
+}
+
+
+
+/// Determine the local index corresponding to VIEW's global index
+/// G_IDX for dimension DIM.
+
+template <typename ViewT>
+index_type
+local_from_global_index(
+ ViewT const& view,
+ dimension_type dim,
+ index_type g_idx)
+VSIP_NOTHROW
+{
+ return view.block().map().impl_local_from_global_index(dim, g_idx);
+}
+
+
+
+/***********************************************************************
+ global_from_local_index
+***********************************************************************/
+
+/// Determine VIEW's global index corresponding to local index L_IDX
+/// of subblock SB.
+
+template <typename ViewT>
+inline
+Index<ViewT::dim>
+global_from_local_index(
+ ViewT const& view,
+ index_type sb,
+ Index<ViewT::dim> const& l_idx)
+{
+ Index<ViewT::dim> g_idx;
+
+ for (dimension_type d=0; d<ViewT::dim; ++d)
+ g_idx[d] = global_from_local_index(view, d, sb, l_idx[d]);
+
+ return g_idx;
+}
+
+
+
+/// Determine VIEW's global index corresponding to local index L_IDX
+/// of the subblock held on the local processor.
+
+template <typename ViewT>
+inline
+Index<ViewT::dim>
+global_from_local_index(
+ ViewT const& view,
+ Index<ViewT::dim> const& l_idx)
+{
+ Index<ViewT::dim> g_idx;
+
+ for (dimension_type d=0; d<ViewT::dim; ++d)
+ g_idx[d] = global_from_local_index(view, d, l_idx[d]);
+
+ return g_idx;
+}
+
+
+
+/// Determine VIEW's global index corresponding to local index L_IDX
+/// for dimension DIM of subblock SB.
+
+template <typename ViewT>
+inline
+index_type
+global_from_local_index(
+ ViewT const& view,
+ dimension_type dim,
+ index_type sb,
+ index_type l_idx)
+{
+ return view.block().map().impl_global_from_local_index(dim, sb, l_idx);
+}
+
+
+
+/// Determine VIEW's global index corresponding to local index L_IDX
+/// for dimension DIM of the subblock held on the local processor.
+
+template <typename ViewT>
+inline
+index_type
+global_from_local_index(
+ ViewT const& view,
+ dimension_type dim,
+ index_type l_idx)
+{
+ return view.block().map().impl_global_from_local_index(dim,
+ view.block().subblock(),
+ l_idx);
+}
+
+} // namespace vsip
+
+#endif // VSIP_IMPL_PAR_SUPPORT_HPP
Index: src/vsip/impl/par-util.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/par-util.hpp,v
retrieving revision 1.7
diff -u -r1.7 par-util.hpp
--- src/vsip/impl/par-util.hpp 11 Jan 2006 16:22:45 -0000 1.7
+++ src/vsip/impl/par-util.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file vsip/impl/par-util.hpp
@author Jules Bergmann
@@ -31,8 +31,6 @@
namespace vsip
{
-const_Vector<processor_type> processor_set();
-
namespace impl
{
@@ -55,25 +53,24 @@
ViewT<T, BlockT> view,
FuncT fcn)
{
- typedef typename BlockT::map_type map_t;
- typedef typename BlockT::local_block_type local_block_t;
- typedef typename map_t::subblock_iterator subblock_iterator;
+ typedef typename BlockT::map_type map_t;
+ typedef typename BlockT::local_block_type local_block_t;
dimension_type const dim = ViewT<T, BlockT>::dim;
BlockT& block = view.block();
- map_t const& am = view.block().map();
+ map_t const& map = view.block().map();
- index_type sb = am.subblock();
+ index_type sb = map.subblock();
if (sb != no_subblock)
{
ViewT<T, local_block_t> local_view = get_local_view(view, sb);
- for (index_type p=0; p<am.num_patches(sb); ++p)
+ for (index_type p=0; p<num_patches(view, sb); ++p)
{
- Domain<dim> ldom = am.template local_domain<dim>(sb, p);
- Domain<dim> gdom = am.template global_domain<dim>(sb, p);
+ Domain<dim> ldom = local_domain(view, sb, p);
+ Domain<dim> gdom = global_domain(view, sb, p);
fcn(local_view.get(ldom), gdom);
}
@@ -82,204 +79,6 @@
-namespace detail
-{
-
-// Return the subdomain of a view/map pair for a subblock.
-
-template <typename ViewT>
-inline Domain<ViewT::dim>
-subblock_domain(
- ViewT const& view,
- Local_map const& /*map*/,
- index_type /*sb*/)
-{
- return block_domain<ViewT::dim>(view.block());
-}
-
-template <typename ViewT,
- typename MapT>
-inline Domain<ViewT::dim>
-subblock_domain(
- ViewT const& /*view*/,
- MapT const& map,
- index_type sb)
-{
- return map.template subblock_domain<ViewT::dim>(sb);
-}
-
-
-
-// Return the local domain of a view/map pair for a subblock/patch.
-
-template <typename ViewT>
-inline Domain<ViewT::dim>
-local_domain(
- ViewT const& view,
- Local_map const& /*map*/,
- index_type /*sb*/,
- index_type /*p*/)
-{
- return block_domain<ViewT::dim>(view.block());
-}
-
-template <typename ViewT,
- typename MapT>
-inline Domain<ViewT::dim>
-local_domain(
- ViewT const& /*view*/,
- MapT const& map,
- index_type sb,
- index_type p)
-{
- return map.template local_domain<ViewT::dim>(sb, p);
-}
-
-
-
-// Return the global domain of a view/map pair for a subblock/patch.
-
-template <typename ViewT>
-inline Domain<ViewT::dim>
-global_domain(
- ViewT const& view,
- Local_map const& /*map*/,
- index_type /*sb*/,
- index_type /*p*/)
-{
- return block_domain<ViewT::dim>(view.block());
-}
-
-template <typename ViewT,
- typename MapT>
-inline Domain<ViewT::dim>
-global_domain(
- ViewT const& /*view*/,
- MapT const& map,
- index_type sb,
- index_type p)
-{
- return map.template global_domain<ViewT::dim>(sb, p);
-}
-
-} // namespace detail
-
-
-
-template <typename ViewT>
-Domain<ViewT::dim>
-subblock_domain(
- ViewT const& view,
- index_type sb)
-{
- return detail::subblock_domain(view, view.block().map(), sb);
-}
-
-template <typename ViewT>
-Domain<ViewT::dim>
-subblock_domain(
- ViewT const& view)
-{
- return detail::subblock_domain(view, view.block().map(),
- view.block().map().subblock());
-}
-
-
-
-template <typename ViewT>
-Domain<ViewT::dim>
-local_domain(
- ViewT const& view,
- index_type sb,
- index_type p)
-{
- return detail::local_domain(view, view.block().map(), sb, p);
-}
-
-template <typename ViewT>
-Domain<ViewT::dim>
-local_domain(
- ViewT const& view,
- index_type p=0)
-{
- return detail::local_domain(view, view.block().map(),
- view.block().map().subblock(),
- p);
-}
-
-
-
-template <typename ViewT>
-Domain<ViewT::dim>
-global_domain(
- ViewT const& view,
- index_type sb,
- index_type p)
-{
- return detail::global_domain(view, view.block().map(), sb, p);
-}
-
-template <typename ViewT>
-Domain<ViewT::dim>
-global_domain(
- ViewT const& view,
- index_type p=0)
-{
- return detail::global_domain(view, view.block().map(),
- view.block().map().subblock(),
- p);
-}
-
-
-
-template <typename ViewT>
-length_type
-num_patches(
- ViewT const& view,
- index_type sb)
-{
- return view.block().map().num_patches(sb);
-}
-
-
-
-template <typename ViewT>
-length_type
-num_patches(
- ViewT const& view)
-{
- return view.block().map().num_patches(view.block().map().subblock());
-}
-
-
-
-template <typename ViewT>
-inline
-index_type
-global_from_local_index(
- ViewT const& view,
- dimension_type dim,
- index_type sb,
- index_type idx)
-{
- return view.block().map().global_from_local_index(dim, sb, idx);
-}
-
-template <typename ViewT>
-inline
-index_type
-global_from_local_index(
- ViewT const& view,
- dimension_type dim,
- index_type idx)
-{
- return view.block().map().global_from_local_index(dim,
- view.block().subblock(),
- idx);
-}
-
-
-
// Evaluate a function object foreach local element of a distributed view.
// Requires:
@@ -309,7 +108,7 @@
{
typename ViewT::local_type local_view = get_local_view(view);
- for (index_type p=0; p<map.num_patches(sb); ++p)
+ for (index_type p=0; p<num_patches(view, sb); ++p)
{
Domain<dim> ldom = local_domain(view, sb, p);
Domain<dim> gdom = global_domain(view, sb, p);
@@ -327,7 +126,6 @@
-
template <typename T, typename Block>
void
buf_send(
Index: src/vsip/impl/replicated_map.hpp
===================================================================
RCS file: src/vsip/impl/replicated_map.hpp
diff -N src/vsip/impl/replicated_map.hpp
--- /dev/null 1 Jan 1970 00:00:00 -0000
+++ src/vsip/impl/replicated_map.hpp 24 Mar 2006 18:55:40 -0000
@@ -0,0 +1,205 @@
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
+
+/** @file vsip/replicated_map.hpp
+ @author Jules Bergmann
+ @date 2005-06-08
+ @brief VSIPL++ Library: Replicated_map class.
+
+*/
+
+#ifndef VSIP_IMPL_REPLICATED_MAP_HPP
+#define VSIP_IMPL_REPLICATED_MAP_HPP
+
+/***********************************************************************
+ Included Files
+***********************************************************************/
+
+#include <vsip/impl/vector-iterator.hpp>
+#include <vsip/impl/par-services.hpp>
+#include <vsip/impl/map-traits.hpp>
+#include <vsip/impl/par-util.hpp>
+#include <vsip/impl/block-traits.hpp>
+#include <vsip/impl/sv_block.hpp>
+
+
+
+/***********************************************************************
+ Declarations
+***********************************************************************/
+
+namespace vsip
+{
+
+template <dimension_type Dim>
+class Replicated_map
+{
+ // Compile-time typedefs.
+public:
+ typedef impl::Vector_iterator<Vector<processor_type> > processor_iterator;
+ typedef std::vector<processor_type> impl_pvec_type;
+ typedef impl::Sv_local_block<processor_type> pset_block_type;
+
+ // Constructor.
+public:
+ Replicated_map()
+ VSIP_THROW((std::bad_alloc));
+
+ template <typename Block>
+ Replicated_map(const_Vector<processor_type, Block> pset)
+ VSIP_THROW((std::bad_alloc));
+
+ // Accessors.
+public:
+ length_type num_subblocks() const VSIP_NOTHROW { return 1; }
+
+ index_type subblock(processor_type pr) const VSIP_NOTHROW
+ {
+ if (this->impl_rank_from_proc(pr) != no_processor)
+ return 0;
+ else
+ return no_subblock;
+ }
+ index_type subblock() const VSIP_NOTHROW
+ { return this->subblock(local_processor()); }
+
+ length_type num_processors() const VSIP_NOTHROW
+ { return this->pset_->size(); }
+
+ processor_iterator processor_begin(index_type sb) const VSIP_NOTHROW
+ {
+ assert(sb == 0);
+ return processor_iterator(this->processor_set(), 0);
+ }
+
+ processor_iterator processor_end (index_type sb) const VSIP_NOTHROW
+ {
+ assert(sb == 0);
+ return processor_iterator(this->processor_set(), this->num_processors());
+ }
+
+ const_Vector<processor_type, pset_block_type> processor_set() const
+ {
+ return const_Vector<processor_type, pset_block_type>(
+ const_cast<pset_block_type&>(*pset_));
+ }
+
+ // Applied map functions.
+public:
+ length_type impl_num_patches(index_type sb) const VSIP_NOTHROW
+ { assert(sb == 0); return 1; }
+
+ void impl_apply(Domain<Dim> const& dom) VSIP_NOTHROW
+ { dom_ = dom; }
+
+ template <dimension_type Dim2>
+ Domain<Dim2> impl_subblock_domain(index_type sb) const VSIP_NOTHROW
+ { assert(sb == 0); return dom_; }
+
+ template <dimension_type Dim2>
+ Domain<Dim2> impl_global_domain(index_type sb, index_type patch)
+ const VSIP_NOTHROW
+ { assert(sb == 0 && patch == 0); return dom_; }
+
+ template <dimension_type Dim2>
+ Domain<Dim2> impl_local_domain (index_type sb, index_type patch)
+ const VSIP_NOTHROW
+ { assert(sb == 0 && patch == 0); return dom_; }
+
+ index_type impl_global_from_local_index(dimension_type /*d*/, index_type sb,
+ index_type idx)
+ const VSIP_NOTHROW
+ { assert(sb == 0); return idx; }
+
+ index_type impl_local_from_global_index(dimension_type /*d*/, index_type idx)
+ const VSIP_NOTHROW
+ { return idx; }
+
+ template <dimension_type Dim2>
+ index_type impl_subblock_from_global_index(Index<Dim2> const& /*idx*/)
+ const VSIP_NOTHROW
+ { return 0; }
+
+ // Extensions.
+public:
+ impl::Communicator impl_comm() const { return impl::default_communicator(); }
+ impl_pvec_type impl_pvec() const
+ { return this->pset_->impl_vector(); }
+
+ length_type impl_working_size() const
+ { return this->num_processors(); }
+
+ processor_type impl_proc_from_rank(index_type idx) const
+ { assert(idx < this->num_processors()); return this->impl_pvec()[idx]; }
+
+ index_type impl_rank_from_proc(processor_type pr) const
+ {
+ for (index_type i=0; i<this->num_processors(); ++i)
+ if (pset_->get(i) == pr) return i;
+ return no_processor;
+ }
+
+ // Member data.
+private:
+ impl::Ref_counted_ptr<pset_block_type> pset_;
+ // impl_pvec_type pset_; // Processor set.
+ Domain<Dim> dom_; // Applied domain.
+};
+
+
+
+template <dimension_type Dim>
+Replicated_map<Dim>::Replicated_map()
+ VSIP_THROW((std::bad_alloc))
+ : pset_(new pset_block_type(vsip::num_processors()), impl::noincrement)
+{
+ for (index_type i=0; i<vsip::num_processors(); ++i)
+ pset_->put(i, vsip::processor_set().get(i));
+}
+
+
+
+// Create a replicated_map with a given processor_set
+//
+// Requires
+// PSET to be a non-empty set of valid processors.
+template <dimension_type Dim>
+template <typename Block>
+Replicated_map<Dim>::Replicated_map(
+ const_Vector<processor_type, Block> pset)
+ VSIP_THROW((std::bad_alloc))
+: pset_(new pset_block_type(pset.size()), impl::noincrement)
+{
+ assert(pset.size() > 0);
+ for (index_type i=0; i<pset.size(); ++i)
+ pset_->put(i, pset.get(i));
+}
+
+
+
+namespace impl
+{
+
+template <dimension_type Dim>
+struct Is_global_map<Replicated_map<Dim> >
+{ static bool const value = true; };
+
+template <dimension_type Dim>
+struct Map_equal<Replicated_map<Dim>, Replicated_map<Dim> >
+{
+ static bool value(Replicated_map<Dim> const& a,
+ Replicated_map<Dim> const& b)
+ {
+ if (a.num_processors() != b.num_processors())
+ return false;
+ for (index_type i=0; i<a.num_processors(); ++i)
+ if (a.impl_proc_from_rank(i) != b.impl_proc_from_rank(i))
+ return false;
+ return true;
+ }
+};
+
+} // namespace vsip::impl
+
+} // namespace vsip
+
+#endif // VSIP_IMPL_REPLICATED_MAP_HPP
Index: src/vsip/impl/subblock.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/src/vsip/impl/subblock.hpp,v
retrieving revision 1.38
diff -u -r1.38 subblock.hpp
--- src/vsip/impl/subblock.hpp 10 Feb 2006 22:24:02 -0000 1.38
+++ src/vsip/impl/subblock.hpp 24 Mar 2006 18:55:40 -0000
@@ -286,7 +286,7 @@
assert(dom_[d].first() < blk_->size(dim, d));
assert(dom_[d].impl_last() < blk_->size(dim, d));
}
- map_.apply(block_domain<dim>(*this));
+ map_.impl_apply(block_domain<dim>(*this));
}
Subset_block(Subset_block const& b)
@@ -294,7 +294,7 @@
dom_ (b.dom_),
map_ (b.map_)
{
- map_.apply(block_domain<dim>(*this));
+ map_.impl_apply(block_domain<dim>(*this));
}
~Subset_block() VSIP_NOTHROW {}
@@ -817,13 +817,13 @@
// Constructors and destructors.
Sliced_block_base(Sliced_block_base const& sb) VSIP_NOTHROW
: map_(sb.map_), blk_(&*sb.blk_), index_(sb.index_)
- { map_.apply(block_domain<dim>(*this)); }
+ { map_.impl_apply(block_domain<dim>(*this)); }
Sliced_block_base(Block &blk, index_type i) VSIP_NOTHROW
: map_(Sliced_block_map<typename Block::map_type,
D>::convert_map(blk.map(), i)),
blk_(&blk),
index_(i)
- { map_.apply(block_domain<dim>(*this)); }
+ { map_.impl_apply(block_domain<dim>(*this)); }
~Sliced_block_base() VSIP_NOTHROW {}
map_type const& map() const { return map_;}
@@ -832,10 +832,10 @@
// The total size of a sliced block is the total size of the underlying
// block, divided by the size of the bound index.
length_type size() const VSIP_NOTHROW
- { return blk_->size() / blk_->size(Block::dim, D);}
+ { return index_ == no_index ? 0 : blk_->size() / blk_->size(Block::dim, D);}
length_type size(dimension_type block_d, dimension_type d) const VSIP_NOTHROW
- { return blk_->size(block_d + 1,
- Compare<dimension_type, D>() > d ? d : d + 1);
+ { return index_ == no_index ? 0 :
+ blk_->size(block_d + 1, Compare<dimension_type, D>() > d ? d : d + 1);
}
// These are noops as Sliced_block is helt by-value.
void increment_count() const VSIP_NOTHROW {}
@@ -1000,13 +1000,13 @@
// Constructors and destructors.
Sliced2_block_base(Sliced2_block_base const& sb) VSIP_NOTHROW
: map_(sb.map_), blk_(&*sb.blk_), index1_(sb.index1_), index2_(sb.index2_)
- { map_.apply(block_domain<dim>(*this)); }
+ { map_.impl_apply(block_domain<dim>(*this)); }
Sliced2_block_base(Block &blk, index_type i, index_type j) VSIP_NOTHROW
: map_(Sliced2_block_map<typename Block::map_type,
D1,
D2>::convert_map(blk.map(), i, j)),
blk_(&blk), index1_(i), index2_(j)
- { map_.apply(block_domain<dim>(*this)); }
+ { map_.impl_apply(block_domain<dim>(*this)); }
~Sliced2_block_base() VSIP_NOTHROW {}
map_type const& map() const { return map_;}
@@ -1277,7 +1277,7 @@
dimension_type const dim = Subset_block<Block>::dim;
- index_type sb = block.map().lookup_index(block.map().impl_rank());
+ index_type sb = block.map().impl_rank_from_proc(local_processor());
Domain<dim> dom = block.impl_block().map().
impl_local_from_global_domain(sb,
@@ -1300,8 +1300,15 @@
typedef Sliced_block<typename Distributed_local_block<Block>::type, D>
local_block_type;
- index_type idx = block.impl_block().map().
- impl_local_from_global_index(D, block.impl_index());
+ // This conversion is only valid if the local processor holds
+ // the subblock containing the slice.
+
+ index_type idx;
+ if (block.map().subblock() != no_subblock)
+ idx = block.impl_block().map().
+ impl_local_from_global_index(D, block.impl_index());
+ else
+ idx = no_index;
return local_block_type(get_local_block(block.impl_block()), idx);
}
Index: src/vsip/impl/sv_block.hpp
===================================================================
RCS file: src/vsip/impl/sv_block.hpp
diff -N src/vsip/impl/sv_block.hpp
--- /dev/null 1 Jan 1970 00:00:00 -0000
+++ src/vsip/impl/sv_block.hpp 24 Mar 2006 18:55:40 -0000
@@ -0,0 +1,114 @@
+/* Copyright (c) 2006 by CodeSourcery, LLC. All rights reserved. */
+
+/** @file vsip/impl/sv_block.hpp
+ @author Jules Bergmann
+ @date 2006-03-09
+ @brief VSIPL++ Library: std::vector block class.
+
+*/
+
+#ifndef VSIP_IMPL_SV_BLOCK_HPP
+#define VSIP_IMPL_SV_BLOCK_HPP
+
+/***********************************************************************
+ Included Files
+***********************************************************************/
+
+#include <vsip/support.hpp>
+#include <vsip/domain.hpp>
+
+#include <vsip/impl/refcount.hpp>
+#include <vsip/impl/block-traits.hpp>
+
+
+
+/***********************************************************************
+ Declarations
+***********************************************************************/
+
+namespace vsip
+{
+
+namespace impl
+{
+
+
+
+template <typename T>
+class Sv_local_block
+ : public impl::Ref_count<Sv_local_block<T> >
+{
+ // Compile-time values and types.
+ typedef std::vector<T> vector_type;
+public:
+ static dimension_type const dim = 1;
+
+ typedef T value_type;
+ typedef T& reference_type;
+ typedef T const& const_reference_type;
+
+ typedef row1_type order_type;
+ typedef Local_map map_type;
+
+ // Implementation types.
+public:
+
+ // Constructors and destructor.
+public:
+ Sv_local_block(Domain<1> const& dom, map_type const& = map_type())
+ VSIP_THROW((std::bad_alloc))
+ : vector_(dom[0].size())
+ {}
+
+ Sv_local_block(Domain<1> const& dom, T value, map_type const& = map_type())
+ VSIP_THROW((std::bad_alloc))
+ : vector_(dom[0].size())
+ {
+ for (index_type i=0; i<dom[0].size(); ++i)
+ this->put(i, value);
+ }
+
+ ~Sv_local_block() VSIP_NOTHROW
+ {}
+
+ // Data accessors.
+public:
+ T get(index_type idx) const VSIP_NOTHROW
+ {
+ assert(idx < size());
+ return vector_[idx];
+ }
+
+ void put(index_type idx, T val) VSIP_NOTHROW
+ {
+ assert(idx < size());
+ vector_[idx] = val;
+ }
+
+ // Accessors.
+public:
+ length_type size() const VSIP_NOTHROW
+ { return vector_.size(); }
+ length_type size(dimension_type D, dimension_type d) const VSIP_NOTHROW
+ { assert(D == 1 && d == 0); return vector_.size(); }
+
+ map_type const& map() const VSIP_NOTHROW { return map_; }
+
+ vector_type impl_vector() { return vector_; }
+
+ // Hidden copy constructor and assignment.
+private:
+ Sv_local_block(Sv_local_block const&);
+ Sv_local_block& operator=(Sv_local_block const&);
+
+ // Member Data
+private:
+ vector_type vector_;
+ map_type map_;
+};
+
+
+} // namespace impl
+} // namespace vsip
+
+#endif // VSIP_IMPL_SV_BLOCK_HPP
Index: tests/appmap.cpp
===================================================================
RCS file: /home/cvs/Repository/vpp/tests/appmap.cpp,v
retrieving revision 1.9
diff -u -r1.9 appmap.cpp
--- tests/appmap.cpp 20 Dec 2005 12:48:40 -0000 1.9
+++ tests/appmap.cpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file tests/appmap.cpp
@author Jules Bergmann
@@ -19,12 +19,22 @@
using namespace std;
using namespace vsip;
+using vsip::impl::Point;
+using vsip::impl::extent_old;
+using vsip::impl::valid;
+using vsip::impl::next;
+using vsip::impl::domain_nth;
+
/***********************************************************************
Definitions
***********************************************************************/
+
+/// Get the nth index in a domain.
+
+
// Utility to create a processor vector of given size.
// Requires:
@@ -70,10 +80,10 @@
if (sb != no_subblock)
{
- for (index_type p=0; p<map.num_patches(sb); ++p)
+ for (index_type p=0; p<map.impl_num_patches(sb); ++p)
{
- Domain<3> gdom = map.template global_domain<3>(sb, p);
- Domain<3> ldom = map.template local_domain<3>(sb, p);
+ Domain<3> gdom = map.template impl_global_domain<3>(sb, p);
+ Domain<3> ldom = map.template impl_local_domain<3>(sb, p);
out << " pr=" << pr << " sb=" << sb << " patch=" << p
<< " gdom=" << gdom
<< " ldom=" << ldom
@@ -85,6 +95,12 @@
+inline Index<1> as_index(Point<1> const& p) {return Index<1>(p[0]); }
+inline Index<2> as_index(Point<2> const& p) {return Index<2>(p[0],p[1]); }
+inline Index<3> as_index(Point<3> const& p) {return Index<3>(p[0],p[1],p[2]); }
+
+
+
// Check that local and global indices within a patch are consistent.
template <dimension_type Dim,
@@ -95,11 +111,26 @@
index_type sb,
index_type p)
{
- Domain<Dim> gdom = map.template global_domain<Dim>(sb, p);
- Domain<Dim> ldom = map.template local_domain<Dim>(sb, p);
+ Domain<Dim> gdom = map.template impl_global_domain<Dim>(sb, p);
+ Domain<Dim> ldom = map.template impl_local_domain<Dim>(sb, p);
test_assert(gdom.size() == ldom.size());
+ length_type dim_num_subblocks[Dim]; // number of subblocks in each dim
+ length_type dim_sb[Dim]; // local sb in each dim
+ length_type dim_num_patches[Dim]; // number of patches in each dim
+ length_type dim_p[Dim]; // local p in each dim
+
+ for (dimension_type d=0; d<Dim; ++d)
+ dim_num_subblocks[d] = map.num_subblocks(d);
+
+ impl::split_tuple(sb, Dim, dim_num_subblocks, dim_sb);
+
+ for (dimension_type d=0; d<Dim; ++d)
+ dim_num_patches[d] = map.impl_subblock_patches(d, dim_sb[d]);
+
+ impl::split_tuple(p, Dim, dim_num_patches, dim_p);
+
for (dimension_type d=0; d<Dim; ++d)
{
test_assert(gdom[d].length() == ldom[d].length());
@@ -109,17 +140,24 @@
index_type gi = gdom[d].impl_nth(i);
index_type li = ldom[d].impl_nth(i);
- if (map.distribution(d) != cyclic)
- {
- test_assert(map.impl_local_from_global_index(d, gi) == li);
- // only valid for 1-dim
- // test_assert(map.impl_subblock_from_index(d, gi) == sb);
- }
- test_assert(map.global_from_local_index(d, sb, li) == gi);
+ test_assert(map.impl_local_from_global_index(d, gi) == li);
+ test_assert(map.impl_global_from_local_index(d, sb, li) == gi);
+ test_assert(map.impl_dim_subblock_from_index(d, gi) == dim_sb[d]);
+ test_assert(map.impl_dim_patch_from_index(d, gi) == dim_p[d]);
}
}
-}
+ Point<Dim> ext = extent_old(gdom);
+
+ for (Point<Dim> idx; valid(ext, idx); next(ext, idx))
+ {
+ Index<Dim> g_idx = as_index(domain_nth(gdom, idx));
+ Index<Dim> l_idx = as_index(domain_nth(ldom, idx));
+
+ test_assert(map.impl_subblock_from_global_index(g_idx) == sb);
+ test_assert(map.impl_patch_from_global_index(g_idx) == p);
+ }
+}
@@ -142,7 +180,7 @@
Vector<processor_type> pvec = create_pvec(num_proc);
map_t map(pvec, dist0);
- map.apply(Domain<dim>(dom.length()));
+ map.impl_apply(Domain<dim>(dom.length()));
Vector<int> data(dom.length(), 0);
@@ -153,9 +191,9 @@
if (sb != no_subblock)
{
- for (index_type p=0; p<map.num_patches(sb); ++p)
+ for (index_type p=0; p<map.impl_num_patches(sb); ++p)
{
- Domain<dim> gdom = map.template global_domain<dim>(sb, p);
+ Domain<dim> gdom = map.template impl_global_domain<dim>(sb, p);
if (gdom.size() > 0)
data(gdom) += 1;
@@ -193,7 +231,7 @@
Vector<processor_type> pvec = create_pvec(num_proc);
map_t map(pvec, dist0, dist1);
- map.apply(dom);
+ map.impl_apply(dom);
Matrix<int> data(dom[0].length(), dom[1].length(), 0);
@@ -204,9 +242,9 @@
if (sb != no_subblock)
{
- for (index_type p=0; p<map.num_patches(sb); ++p)
+ for (index_type p=0; p<map.impl_num_patches(sb); ++p)
{
- Domain<dim> gdom = map.template global_domain<dim>(sb, p);
+ Domain<dim> gdom = map.template impl_global_domain<dim>(sb, p);
data(gdom) += 1;
@@ -294,23 +332,26 @@
Vector<processor_type> pvec = create_pvec(num_proc);
map_t map(pvec, Block_dist(4), Block_dist(4));
- map.apply(Domain<3>(16, 16, 1));
+ map.impl_apply(Domain<3>(16, 16, 1));
- test_assert(map.num_patches(0) == 1);
- test_assert(map.global_domain<3>(0, 0) == Domain<3>(Domain<1>(0, 1, 4),
- Domain<1>(0, 1, 4),
- Domain<1>(0, 1, 1)));
+ test_assert(map.impl_num_patches(0) == 1);
+ test_assert(map.impl_global_domain<3>(0, 0) ==
+ Domain<3>(Domain<1>(0, 1, 4),
+ Domain<1>(0, 1, 4),
+ Domain<1>(0, 1, 1)));
// subblocks are row-major
- test_assert(map.num_patches(1) == 1);
- test_assert(map.global_domain<3>(1, 0) == Domain<3>(Domain<1>(0, 1, 4),
- Domain<1>(4, 1, 4),
- Domain<1>(0, 1, 1)));
-
- test_assert(map.num_patches(15) == 1);
- test_assert(map.global_domain<3>(15, 0) == Domain<3>(Domain<1>(12, 1, 4),
- Domain<1>(12, 1, 4),
- Domain<1>(0, 1, 1)));
+ test_assert(map.impl_num_patches(1) == 1);
+ test_assert(map.impl_global_domain<3>(1, 0) ==
+ Domain<3>(Domain<1>(0, 1, 4),
+ Domain<1>(4, 1, 4),
+ Domain<1>(0, 1, 1)));
+
+ test_assert(map.impl_num_patches(15) == 1);
+ test_assert(map.impl_global_domain<3>(15, 0) ==
+ Domain<3>(Domain<1>(12, 1, 4),
+ Domain<1>(12, 1, 4),
+ Domain<1>(0, 1, 1)));
}
Index: tests/replicated_map.cpp
===================================================================
RCS file: tests/replicated_map.cpp
diff -N tests/replicated_map.cpp
--- /dev/null 1 Jan 1970 00:00:00 -0000
+++ tests/replicated_map.cpp 24 Mar 2006 18:55:40 -0000
@@ -0,0 +1,124 @@
+/* Copyright (c) 2006 by CodeSourcery, LLC. All rights reserved. */
+
+/** @file tests/replicated_map.cpp
+ @author Jules Bergmann
+ @date 2006-03-09
+ @brief VSIPL++ Library: Unit tests for Replicated_map.
+*/
+
+/***********************************************************************
+ Included Files
+***********************************************************************/
+
+#include <vsip/support.hpp>
+#include <vsip/map.hpp>
+#include <vsip/initfin.hpp>
+
+#include "test.hpp"
+
+using namespace std;
+using namespace vsip;
+
+
+
+/***********************************************************************
+ Definitions
+***********************************************************************/
+
+template <dimension_type Dim,
+ typename Block>
+void
+check_replicated_map(
+ Replicated_map<Dim> const& map,
+ const_Vector<processor_type, Block> pset)
+{
+ typedef Replicated_map<Dim> map_type;
+ typedef typename map_type::processor_iterator iterator;
+
+ // Check num_processors()
+ test_assert(map.num_processors() == pset.size());
+
+ // Check processor_set()
+ Vector<processor_type> map_pset = map.processor_set();
+
+ test_assert(map_pset.size() == pset.size());
+ for (index_type i=0; i<map_pset.size(); ++i)
+ test_assert(map_pset(i) == pset(i));
+
+ // Check processor_begin(), processor_end()
+ iterator begin = map.processor_begin(0);
+ iterator end = map.processor_end(0);
+
+ assert(static_cast<length_type>(end - begin) == pset.size());
+
+ iterator cur = begin;
+ while (cur != end)
+ {
+ index_type i = cur - begin;
+ test_assert(*cur == pset(i));
+ ++cur;
+ }
+}
+
+
+
+// Check that map can be constructed with a processor set.
+
+template <dimension_type Dim,
+ typename Block>
+void
+test_single_pset(const_Vector<processor_type, Block> pset)
+{
+ typedef Replicated_map<Dim> map_type;
+
+ map_type map(pset);
+ check_replicated_map(map, pset);
+}
+
+
+
+// Check that map can be constructed with default processor set.
+
+template <dimension_type Dim>
+void
+test_default_pset()
+{
+ typedef Replicated_map<Dim> map_type;
+
+ map_type map;
+ check_replicated_map(map, vsip::processor_set());
+}
+
+
+
+template <dimension_type Dim>
+void
+test_pset()
+{
+ Vector<processor_type> vec1(1);
+ Vector<processor_type> vec4(4);
+
+ vec1(0) = 1;
+
+ vec4(0) = 3;
+ vec4(1) = 2;
+ vec4(2) = 1;
+ vec4(3) = 0;
+
+ test_single_pset<Dim>(vec1);
+ test_single_pset<Dim>(vec4);
+
+ test_default_pset<Dim>();
+}
+
+
+
+int
+main(int argc, char** argv)
+{
+ vsipl vpp(argc, argv);
+
+ test_pset<1>();
+ test_pset<2>();
+ test_pset<3>();
+}
Index: tests/util-par.hpp
===================================================================
RCS file: /home/cvs/Repository/vpp/tests/util-par.hpp,v
retrieving revision 1.7
diff -u -r1.7 util-par.hpp
--- tests/util-par.hpp 5 Dec 2005 19:19:19 -0000 1.7
+++ tests/util-par.hpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file tests/util-par.hpp
@author Jules Bergmann
@@ -19,6 +19,7 @@
#include <vsip/vector.hpp>
#include <vsip/matrix.hpp>
+#include <vsip/parallel.hpp>
#include "output.hpp"
#include "extdata-output.hpp"
@@ -123,10 +124,11 @@
map_t const& am = view.block().map();
msg(am, std::string(name) + " ------------------------------------------\n");
- std::cout << "(" << am.impl_rank() << "): dump_view(" << name << ")\n";
- std::cout << "(" << am.impl_rank() << "): map "
+ std::cout << "(" << vsip::local_processor() << "): dump_view(" << name
+ << ")\n";
+ std::cout << "(" << vsip::local_processor() << "): map "
<< Type_name<map_t>::name() << "\n";
- std::cout << "(" << am.impl_rank() << "): block "
+ std::cout << "(" << vsip::local_processor() << "): block "
<< Type_name<Block>::name() << "\n";
index_type sb = am.subblock();
@@ -181,7 +183,8 @@
map_t const& am = view.block().map();
msg(am, std::string(name) + " ------------------------------------------\n");
- std::cout << "(" << am.impl_rank() << "): dump_view(Matrix " << name << ")\n";
+ std::cout << "(" << vsip::local_processor() << "): dump_view(Matrix "
+ << name << ")\n";
index_type sb = am.subblock();
if (sb != no_subblock)
@@ -209,7 +212,7 @@
index_type gr = gdom[0].impl_nth(r);
index_type gc = gdom[1].impl_nth(c);
- std::cout << "(" << am.impl_rank() << ") " << sb << "/" << p
+ std::cout << "(" << vsip::local_processor() << ") " << sb << "/" << p
<< " ["
<< lr << "," << lc << ":"
<< gr << "," << gc << "] = "
@@ -234,12 +237,12 @@
dump_map(MapT const& map)
{
typedef typename MapT::processor_iterator p_iter_t;
- vsip::processor_type rank = map.impl_rank();
+ vsip::processor_type rank = vsip::local_processor();
std::ostringstream s;
- s << map.impl_proc(0);
- for (vsip::index_type i=1; i<map.impl_num_procs(); ++i)
- s << "," << map.impl_proc(i);
+ s << map.impl_proc_from_rank(0);
+ for (vsip::index_type i=1; i<map.num_processors(); ++i)
+ s << "," << map.impl_proc_from_rank(i);
std::cout << rank << ": " << Type_name<MapT>::name()
<< " [" << s.str() << "]"
@@ -248,8 +251,8 @@
for (vsip::index_type sb=0; sb<map.num_subblocks(); ++sb)
{
std::cout << " sub " << sb << ": ";
- if (map.impl_is_applied())
- std::cout << map.template global_domain<Dim>(sb, 0);
+ // if (map.impl_is_applied())
+ std::cout << map.template impl_global_domain<Dim>(sb, 0);
std::cout << " [";
for (p_iter_t p=map.processor_begin(sb); p != map.processor_end(sb); ++p)
Index: tests/parallel/block.cpp
===================================================================
RCS file: /home/cvs/Repository/vpp/tests/parallel/block.cpp,v
retrieving revision 1.1
diff -u -r1.1 block.cpp
--- tests/parallel/block.cpp 16 Mar 2006 03:27:10 -0000 1.1
+++ tests/parallel/block.cpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file tests/parallel/block.cpp
@author Jules Bergmann
@@ -58,17 +58,14 @@
typename ViewT<T, BlockT>::local_type lview = view.local();
index_type sb = map.subblock();
+
+ Domain<Dim> dom = subblock_domain(view, sb);
+ test_assert(lview.size() == impl::size(dom));
+ for (dimension_type d=0; d<Dim; ++d)
+ test_assert(lview.size(d) == dom[d].size());
+
if (sb == no_subblock)
- {
test_assert(lview.size() == 0);
- }
- else
- {
- Domain<Dim> dom = map.template subblock_domain<Dim>(sb);
- test_assert(lview.size() == impl::size(dom));
- for (dimension_type d=0; d<Dim; ++d)
- test_assert(lview.size(d) == dom[d].size());
- }
}
@@ -161,7 +158,7 @@
Par_assign<Dim, T, T, dist_block0_t, dist_block2_t>
a3(view0, view2);
- // cout << "(" << map1.impl_rank() << "): test_distributed_view\n";
+ // cout << "(" << local_processor() << "): test_distributed_view\n";
for (int l=0; l<loop; ++l)
{
@@ -183,7 +180,7 @@
typename view0_t::local_type local_view = view0.local();
- if (map1.impl_rank() == 0)
+ if (local_processor() == 0)
{
// On processor 0, local_view should be entire view.
test_assert(extent_old(local_view) == extent_old(dom));
@@ -262,7 +259,7 @@
impl::Communicator comm = impl::default_communicator();
- // cout << "(" << map1.impl_rank() << "): test_distributed_view\n";
+ // cout << "(" << local_processor() << "): test_distributed_view\n";
for (int l=0; l<loop; ++l)
{
@@ -282,7 +279,7 @@
// Check results.
comm.barrier();
- if (map1.impl_rank() == 0)
+ if (local_processor() == 0)
{
typename view0_t::local_type local_view = view0.local();
Index: tests/parallel/corner-turn.cpp
===================================================================
RCS file: tests/parallel/corner-turn.cpp
diff -N tests/parallel/corner-turn.cpp
--- /dev/null 1 Jan 1970 00:00:00 -0000
+++ tests/parallel/corner-turn.cpp 24 Mar 2006 18:55:40 -0000
@@ -0,0 +1,98 @@
+/* Copyright (c) 2006 by CodeSourcery, LLC. All rights reserved. */
+
+/** @file tests/corner-turn.cpp
+ @author Jules Bergmann
+ @date 2006-02-14
+ @brief VSIPL++ Library: Functional test for corner-turns.
+*/
+
+/***********************************************************************
+ Included Files
+***********************************************************************/
+
+#include <iostream>
+#include <unistd.h>
+
+#include <vsip/initfin.hpp>
+#include <vsip/support.hpp>
+#include <vsip/map.hpp>
+#include <vsip/tensor.hpp>
+#include <vsip/parallel.hpp>
+
+#include "test.hpp"
+#include "output.hpp"
+#include "util.hpp"
+#include "util-par.hpp"
+
+using namespace std;
+using namespace vsip;
+
+
+
+/***********************************************************************
+ Definitions
+***********************************************************************/
+
+int
+main(int argc, char** argv)
+{
+ vsipl vpp(argc, argv);
+
+#if 0
+ // Enable this section for easier debugging.
+ impl::Communicator comm = impl::default_communicator();
+ pid_t pid = getpid();
+
+ cout << "rank: " << comm.rank()
+ << " size: " << comm.size()
+ << " pid: " << pid
+ << endl;
+
+ // Stop each process, allow debugger to be attached.
+ if (comm.rank() == 0) fgetc(stdin);
+ comm.barrier();
+ cout << "start\n";
+#endif
+
+ typedef float T;
+
+ typedef Map<Block_dist, Block_dist> map_type;
+ typedef Dense<2, T, row2_type, map_type> block_type;
+
+ processor_type np = num_processors();
+
+ length_type rows = 32;
+ length_type cols = 64;
+
+ map_type root_map(1, 1);
+ map_type row_map (np, 1);
+ map_type col_map (1, np);
+
+ Matrix<T, block_type> src(rows, cols, root_map);
+ Matrix<T, block_type> A (rows, cols, row_map);
+ Matrix<T, block_type> B (rows, cols, col_map);
+ Matrix<T, block_type> dst(rows, cols, root_map);
+
+ if (root_map.subblock() != no_subblock)
+ {
+ // cout << local_processor() << "/" << np << ": initializing " << endl;
+ for (index_type r=0; r<rows; ++r)
+ for (index_type c=0; c<rows; ++c)
+ src.local().put(r, c, T(r*cols+c));
+ }
+
+ A = src; // scatter
+ B = A; // corner-turn
+ dst = B; // gather
+
+ if (root_map.subblock() != no_subblock)
+ {
+ // cout << local_processor() << "/" << np << ": checking " << endl;
+ for (index_type r=0; r<rows; ++r)
+ for (index_type c=0; c<rows; ++c)
+ test_assert(equal(src.local().get(r, c),
+ dst.local().get(r, c)));
+ }
+
+ return 0;
+}
Index: tests/parallel/expr.cpp
===================================================================
RCS file: /home/cvs/Repository/vpp/tests/parallel/expr.cpp,v
retrieving revision 1.1
diff -u -r1.1 expr.cpp
--- tests/parallel/expr.cpp 16 Mar 2006 03:27:10 -0000 1.1
+++ tests/parallel/expr.cpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file tests/parallel/expr.cpp
@author Jules Bergmann
@@ -159,7 +159,7 @@
impl::Communicator comm = impl::default_communicator();
- // cout << "(" << map1.impl_rank() << "): test_distributed_view\n";
+ // cout << "(" << local_processor() << "): test_distributed_view\n";
foreach_point(A, Set_identity<Dim>(dom, 2, 1));
foreach_point(B, Set_identity<Dim>(dom, 3, 2));
@@ -213,7 +213,7 @@
foreach_point(Z7, checker5);
test_assert(checker5.good());
- if (map_res.impl_rank() == 0) // rank(map_res) == 0
+ if (map_res.subblock() != no_subblock)
{
typename view0_t::local_type local_view = chk1.local();
@@ -325,7 +325,7 @@
// Check results.
comm.barrier();
- if (map_res.impl_rank() == 0)
+ if (map0.subblock() != no_subblock)
{
typename view0_t::local_type local_view = chk1.local();
@@ -434,7 +434,7 @@
// Check results.
comm.barrier();
- if (map_res.impl_rank() == 0)
+ if (map0.subblock() != no_subblock)
{
typename view0_t::local_type local_view = chk.local();
@@ -636,6 +636,13 @@
Global_map<1>(),
loop);
+ test_distributed_expr<float>(
+ Domain<1>(16),
+ Replicated_map<1>(),
+ Replicated_map<1>(),
+ Replicated_map<1>(),
+ loop);
+
test_vector_assign<float>(loop);
test_matrix_assign<float>(loop);
Index: tests/parallel/fftm.cpp
===================================================================
RCS file: /home/cvs/Repository/vpp/tests/parallel/fftm.cpp,v
retrieving revision 1.1
diff -u -r1.1 fftm.cpp
--- tests/parallel/fftm.cpp 16 Mar 2006 03:27:10 -0000 1.1
+++ tests/parallel/fftm.cpp 24 Mar 2006 18:55:40 -0000
@@ -1,4 +1,4 @@
-/* Copyright (c) 2005 by CodeSourcery, LLC. All rights reserved. */
+/* Copyright (c) 2005, 2006 by CodeSourcery, LLC. All rights reserved. */
/** @file tests/parallel/fftm.cpp
@author Nathan Myers
@@ -101,51 +101,23 @@
test_assert(in.size(0) == 5);
length_type const N = in.size(1);
- Block& block = in.block();
- Matrix<T,Dense<2,T,tuple<0,1,2>,Local_map> > local_in(
- block.get_local_block());
+ in.row(0) = T();
- if (block.subblock() == block.map().impl_subblock_from_index(0,0))
- {
- unsigned ix = block.map().impl_local_from_global_index(0,0);
- local_in.row(ix) = T();
- }
-
- if (block.subblock() == block.map().impl_subblock_from_index(0,1))
- {
- unsigned ix = block.map().impl_local_from_global_index(0,1);
- local_in.row(ix) = T();
- local_in.row(ix)(0) = T(scale);
- }
+ in.row(1) = T();
+ in.row(1)(0) = T(scale);
- if (block.subblock() == block.map().impl_subblock_from_index(0,2))
- {
- unsigned ix = block.map().impl_local_from_global_index(0,2);
- local_in.row(ix) = T();
- local_in.row(ix)(0) = T(1);
- local_in.row(ix)(Domain<1>(0, 1, N)) += T(3);
- if (local_in.size(1) > 4)
- local_in.row(ix)(Domain<1>(0, 4, N/4)) += T(-2);
- if (local_in.size(1) > 13)
- local_in.row(ix)(Domain<1>(0, 13, N/13)) += T(7);
- if (local_in.size(1) > 27)
- local_in.row(ix)(Domain<1>(0, 27, N/27)) += T(-15);
- if (local_in.size(1) > 37)
- local_in.row(ix)(Domain<1>(0, 37, N/37)) += T(31);
- }
+ in.row(2) = T();
+ in.row(2)(0) = T(1);
+ in.row(2)(Domain<1>(0, 1, N)) += T(3);
+ if (in.size(1) > 4) in.row(2)(Domain<1>(0, 4, N/4)) += T(-2);
+ if (in.size(1) > 13) in.row(2)(Domain<1>(0, 13, N/13)) += T(7);
+ if (in.size(1) > 27) in.row(2)(Domain<1>(0, 27, N/27)) += T(-15);
+ if (in.size(1) > 37) in.row(2)(Domain<1>(0, 37, N/37)) += T(31);
- if (block.subblock() == block.map().impl_subblock_from_index(0,3))
- {
- unsigned ix = block.map().impl_local_from_global_index(0,3);
- local_in.row(ix) = T(scale);
- }
+ in.row(3) = T(scale);
- if (block.subblock() == block.map().impl_subblock_from_index(0,4))
- {
- unsigned ix = block.map().impl_local_from_global_index(0,4);
- for (unsigned i = 0; i < N; ++i)
- local_in.row(ix)(i) = T(std::sin(3.1415926535898*i*4/N));
- }
+ for (unsigned i = 0; i < N; ++i)
+ in.row(4)(i) = T(std::sin(3.1415926535898*i*4/N));
}
@@ -269,96 +241,23 @@
test_assert(in.size(1) == 5);
length_type const N = in.size(0);
-#if 0
- typedef Map<Block_dist,Block_dist> map_type;
- map_type map(Block_dist(1), Block_dist(1));
-
- typedef Dense<2,T,tuple<0,1,2>,map_type> here_block_type;
- typedef Matrix<T,here_block_type> here_matrix_type;
-
- Domain<2> domain(Domain<1>(N), Domain<1>(5));
- here_block_type in_block(domain, map);
- here_matrix_type in(in_block);
-#endif
-
- Block& block = in.block();
- Matrix<T,Dense<2,T,tuple<0,1,2>,Local_map> > local_in(
- block.get_local_block());
-
-#if 1
- if (block.subblock() == block.map().impl_subblock_from_index(1,0))
- {
- unsigned ix = block.map().impl_local_from_global_index(1,0);
- local_in.col(ix) = T();
- }
-
- if (block.subblock() == block.map().impl_subblock_from_index(1,1))
- {
- unsigned ix = block.map().impl_local_from_global_index(1,1);
- local_in.col(ix) = T();
- local_in.col(ix)(0) = T(scale);
- }
-
- if (block.subblock() == block.map().impl_subblock_from_index(1,2))
- {
- unsigned ix = block.map().impl_local_from_global_index(1,2);
- local_in.col(ix) = T();
- local_in.col(ix)(0) = T(1);
- local_in.col(ix)(Domain<1>(0, 1, N)) += T(3);
- if (local_in.size(0) > 4)
- local_in.col(ix)(Domain<1>(0, 4, N/4)) += T(-2);
- if (local_in.size(0) > 13)
- local_in.col(ix)(Domain<1>(0, 13, N/13)) += T(7);
- if (local_in.size(0) > 27)
- local_in.col(ix)(Domain<1>(0, 27, N/27)) += T(-15);
- if (local_in.size(0) > 37)
- local_in.col(ix)(Domain<1>(0, 37, N/37)) += T(31);
- }
+ in.col(0) = T();
- if (block.subblock() == block.map().impl_subblock_from_index(1,3))
- {
- unsigned ix = block.map().impl_local_from_global_index(1,3);
- local_in.col(ix) = T(scale);
- }
+ in.col(1) = T();
+ in.col(1)(0) = T(scale);
- if (block.subblock() == block.map().impl_subblock_from_index(1,4))
- {
- unsigned ix = block.map().impl_local_from_global_index(1,4);
- for (unsigned i = 0; i < N; ++i)
- local_in.col(ix)(i) = T(std::sin(3.1415926535898*i*4/N));
- }
-
-#endif
+ in.col(2) = T();
+ in.col(2)(0) = T(1);
+ in.col(2)(Domain<1>(0, 1, N)) += T(3);
+ if (in.size(0) > 4) in.col(2)(Domain<1>(0, 4, N/4)) += T(-2);
+ if (in.size(0) > 13) in.col(2)(Domain<1>(0, 13, N/13)) += T(7);
+ if (in.size(0) > 27) in.col(2)(Domain<1>(0, 27, N/27)) += T(-15);
+ if (in.size(0) > 37) in.col(2)(Domain<1>(0, 37, N/37)) += T(31);
-#if 0
- impl::Communicator comm = impl::default_communicator();
- if (comm.rank() == 0)
- {
- local_in.col(0) = T();
-
- local_in.col(1) = T();
- local_in.col(1)(0) = T(scale);
-
- local_in.col(2) = T();
- local_in.col(2)(0) = T(1);
- local_in.col(2)(Domain<1>(0, 1, N)) += T(3);
- local_in.col(2)(Domain<1>(0, 4, N/4)) += T(-2);
- if (in.size(0) > 16)
- local_in.col(2)(Domain<1>(0, 13, N/13)) += T(7);
- if (in.size(0) > 27)
- local_in.col(2)(Domain<1>(0, 27, N/27)) += T(-15);
- if (in.size(0) > 37)
- local_in.col(2)(Domain<1>(0, 37, N/37)) += T(31);
+ in.col(3) = T(scale);
- local_in.col(3) = T(scale);
-
- for (unsigned i = 0; i < N; ++i)
- local_in.col(4)(i) = T(std::sin(3.1415926535898*i*4/N));
-
- inp = in;
- }
- comm.barrier();
-#endif
+ for (unsigned i = 0; i < N; ++i)
+ in.col(4)(i) = T(std::sin(3.1415926535898*i*4/N));
}
@@ -592,23 +491,11 @@
vsipl init(argc, argv);
- impl::Communicator comm = impl::default_communicator();
- ::number_of_processors = comm.size();
+ ::number_of_processors = vsip::num_processors();
-#if 0
- // Enable this section for easier debugging.
- pid_t pid = getpid();
-
- cout << "rank: " << comm.rank()
- << " size: " << comm.size()
- << " pid: " << pid
- << endl;
-
- // Stop each process, allow debugger to be attached.
- char c;
- if (comm.rank() == 0) read(0,&c,1);
- comm.barrier();
-#endif
+ // include debug.hpp to use this function. It waits for input
+ // to allow a debugger to be attached.
+ // debug_stub();
#if defined(VSIP_IMPL_FFT_USE_FLOAT)
test<float>();
Index: tests/parallel/replicated_data.cpp
===================================================================
RCS file: tests/parallel/replicated_data.cpp
diff -N tests/parallel/replicated_data.cpp
--- /dev/null 1 Jan 1970 00:00:00 -0000
+++ tests/parallel/replicated_data.cpp 24 Mar 2006 18:55:40 -0000
@@ -0,0 +1,338 @@
+/* Copyright (c) 2006 by CodeSourcery, LLC. All rights reserved. */
+
+/** @file tests/parallel/replicated_data.cpp
+ @author Jules Bergmann
+ @date 2006-03-10
+ @brief VSIPL++ Library: Unit tests for using Replicated_map.
+*/
+
+/***********************************************************************
+ Included Files
+***********************************************************************/
+
+#define VERBOSE 0
+
+#if VERBOSE
+# include <iostream>
+#endif
+#include <algorithm>
+
+#include <vsip/support.hpp>
+#include <vsip/map.hpp>
+#include <vsip/initfin.hpp>
+
+#include "test.hpp"
+#include "output.hpp"
+#include "debug.hpp"
+
+using namespace std;
+using namespace vsip;
+
+
+
+/***********************************************************************
+ Definitions
+***********************************************************************/
+
+void msg(char* text)
+{
+ impl::default_communicator().barrier();
+ if (local_processor() == 0)
+ std::cout << text << std::endl;
+ impl::default_communicator().barrier();
+}
+
+
+
+// Test comm using replicated view as source (to non-replicated view)
+
+template <typename T>
+void
+test_src(int modulo)
+{
+ // rep_view will be replicated over a subset of processors.
+ // dst_view will dim 0 distributed so that each processor has a row.
+
+ length_type np = num_processors();
+
+ Vector<processor_type> full_pset = processor_set();
+ Vector<processor_type> pset(1+(np-1)/modulo);
+
+ for (index_type i=0; i<np; i+=modulo)
+ pset(i/modulo) = full_pset(i);
+
+ length_type size = 16;
+ length_type rows = np;
+
+ Replicated_map<1> rep_map(pset);
+ Map<> dst_map(np, 1);
+
+ typedef Dense<1, T, row1_type, Replicated_map<1> > rep_block_type;
+ typedef Dense<2, T, row2_type, Map<> > dst_block_type;
+
+ typedef Vector<T, rep_block_type> rep_view_type;
+ typedef Matrix<T, dst_block_type> dst_view_type;
+
+ rep_view_type rep_view(size, rep_map);
+ dst_view_type dst_view(rows, size, dst_map);
+
+ for (index_type i=0; i<size; ++i)
+ rep_view.put(i, T(i));
+
+ if (rep_map.subblock() != no_subblock)
+ {
+ rep_view.local().put(0, T(local_processor()));
+ }
+
+ for (index_type r=0; r<rows; ++r)
+ dst_view.row(r) = rep_view;
+
+ // Check the results
+ typename dst_view_type::local_type l_dst_view = dst_view.local();
+ test_assert(l_dst_view.size(0) == 1);
+
+ for (index_type i=1; i<size; ++i)
+ test_assert(l_dst_view.get(0, i) == T(i));
+
+#if VERBOSE
+ std::cout << "(" << local_processor() << "): " << l_dst_view.get(0, 0)
+ << std::endl;
+#endif
+}
+
+
+
+// Test comm using replicated view as destination (from non-replicated view)
+
+template <typename T,
+ typename SrcMapT,
+ typename DstMapT>
+void
+test_msg(
+ length_type size,
+ SrcMapT const& src_map,
+ DstMapT const& dst_map,
+ bool mark_first_element = true)
+{
+ typedef Dense<1, T, row1_type, SrcMapT> src_block_type;
+ typedef Dense<1, T, row1_type, DstMapT> dst_block_type;
+
+ typedef Vector<T, src_block_type> src_view_type;
+ typedef Vector<T, dst_block_type> dst_view_type;
+
+ src_view_type src_view(size, src_map);
+ dst_view_type dst_view(size, dst_map);
+
+ for (index_type i=0; i<size; ++i)
+ src_view.put(i, T(i));
+
+ if (mark_first_element && src_map.subblock() != no_subblock)
+ {
+ src_view.local().put(0, T(local_processor()));
+ }
+
+ dst_view = src_view;
+
+ // Check the results
+ if (dst_map.subblock() != no_subblock)
+ {
+ typename dst_view_type::local_type l_dst_view = dst_view.local();
+
+ for (index_type li=1; li<l_dst_view.size(); ++li)
+