From a680e329e62ac1fba9e7512e3da66623ceeee8ba Mon Sep 17 00:00:00 2001
From: Wenkai Du <43822138+wenkaidu@users.noreply.github.com>
Date: Sat, 12 Oct 2024 08:43:16 -0700
Subject: [PATCH] Temporarily disable MSCCL all gather XMLs due to UT failure
 (#1373)

---
 tools/msccl-algorithms/allgather-8n-0-8kb.xml |  356 ----
 .../allgather-8n-128kb-1mb.xml                | 1364 -------------
 .../allgather-8n-1mb-40mb.xml                 | 1812 -----------------
 .../allgather-8n-8kb-128kb.xml                |  692 -------
 .../allgather-allpairs-16n-16tb.xml           |  994 ---------
 5 files changed, 5218 deletions(-)
 delete mode 100644 tools/msccl-algorithms/allgather-8n-0-8kb.xml
 delete mode 100644 tools/msccl-algorithms/allgather-8n-128kb-1mb.xml
 delete mode 100644 tools/msccl-algorithms/allgather-8n-1mb-40mb.xml
 delete mode 100644 tools/msccl-algorithms/allgather-8n-8kb-128kb.xml
 delete mode 100644 tools/msccl-algorithms/allgather-allpairs-16n-16tb.xml

diff --git a/tools/msccl-algorithms/allgather-8n-0-8kb.xml b/tools/msccl-algorithms/allgather-8n-0-8kb.xml
deleted file mode 100644
index 6b43e13b4d..0000000000
--- a/tools/msccl-algorithms/allgather-8n-0-8kb.xml
+++ /dev/null
@@ -1,356 +0,0 @@
-<!-- Copyright (c) Microsoft Corporation. -->
-<!-- Licensed under the MIT License. -->
-<algo name="all_gather_llm" proto="LL" nchannels="1" nchunksperloop="8" ngpus="8" coll="allgather" inplace="1" outofplace="0" minBytes="0" maxBytes="8192">
-  <gpu id="0" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="1" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="2" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="3" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="4" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="5" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="6" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="7" i_chunks="0" o_chunks="8" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-</algo>
diff --git a/tools/msccl-algorithms/allgather-8n-128kb-1mb.xml b/tools/msccl-algorithms/allgather-8n-128kb-1mb.xml
deleted file mode 100644
index 03e14f2feb..0000000000
--- a/tools/msccl-algorithms/allgather-8n-128kb-1mb.xml
+++ /dev/null
@@ -1,1364 +0,0 @@
-<!-- Copyright (c) Microsoft Corporation. -->
-<!-- Licensed under the MIT License. -->
-<algo name="all_gather_llm" proto="Simple" nchannels="32" nchunksperloop="32" ngpus="8" coll="allgather" inplace="1" outofplace="0" minBytes="131073" maxBytes="1048575">
-  <gpu id="0" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="1" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="1" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="2" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="2" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="3" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="3" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="4" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="4" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="16">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="17">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="1" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="2" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="2" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="2" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="2" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="3" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="3" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="3" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="3" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="4" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="4" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="4" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="4" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="8">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="9">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="24">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="25">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="2" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="1" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="1" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="1" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="1" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="3" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="3" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="3" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="3" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="4" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="4" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="4" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="4" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="18">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="19">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="3" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="1" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="1" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="1" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="1" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="2" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="2" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="2" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="2" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="4" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="4" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="4" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="4" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="10">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="11">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="26">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="27">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="4" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="1" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="1" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="1" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="1" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="2" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="2" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="2" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="2" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="3" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="3" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="3" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="3" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="20">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="21">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="5" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="1" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="1" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="1" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="1" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="2" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="2" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="2" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="2" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="3" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="3" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="3" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="3" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="4" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="4" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="4" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="4" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="12">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="13">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="28">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="29">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="6" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="7" chan="14">
-      <step s="0" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="7" chan="15">
-      <step s="0" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="7" chan="30">
-      <step s="0" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="7" chan="31">
-      <step s="0" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="1" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="1" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="1" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="1" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="2" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="2" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="2" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="2" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="3" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="3" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="3" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="3" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="4" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="4" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="4" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="4" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="5" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="5" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="5" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="5" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="-1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="-1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="-1" chan="22">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="-1" chan="23">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="7" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="0" chan="16">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="0" chan="17">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="1" chan="8">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="1" chan="9">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="1" chan="24">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="1" chan="25">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="2" chan="2">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="2" chan="3">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="2" chan="18">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="2" chan="19">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="3" chan="10">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="3" chan="11">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="-1" recv="3" chan="26">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="-1" recv="3" chan="27">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="-1" recv="4" chan="4">
-      <step s="0" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="-1" recv="4" chan="5">
-      <step s="0" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="-1" recv="4" chan="20">
-      <step s="0" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="-1" recv="4" chan="21">
-      <step s="0" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="-1" recv="5" chan="12">
-      <step s="0" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="-1" recv="5" chan="13">
-      <step s="0" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="-1" recv="5" chan="28">
-      <step s="0" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="-1" recv="5" chan="29">
-      <step s="0" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="-1" recv="6" chan="6">
-      <step s="0" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="-1" recv="6" chan="7">
-      <step s="0" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="-1" recv="6" chan="22">
-      <step s="0" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="-1" recv="6" chan="23">
-      <step s="0" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="0" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="0" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="0" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="0" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="1" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="1" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="1" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="1" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="2" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="2" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="2" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="2" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="3" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="3" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="3" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="3" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="4" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="4" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="4" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="4" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="5" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="5" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="5" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="5" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="6" recv="-1" chan="14">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="6" recv="-1" chan="15">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="6" recv="-1" chan="30">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="6" recv="-1" chan="31">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-</algo>
diff --git a/tools/msccl-algorithms/allgather-8n-1mb-40mb.xml b/tools/msccl-algorithms/allgather-8n-1mb-40mb.xml
deleted file mode 100644
index da1de14d6c..0000000000
--- a/tools/msccl-algorithms/allgather-8n-1mb-40mb.xml
+++ /dev/null
@@ -1,1812 +0,0 @@
-<!-- Copyright (c) Microsoft Corporation. -->
-<!-- Licensed under the MIT License. -->
-<algo name="all_gather_llm" proto="Simple" nchannels="8" nchunksperloop="64" ngpus="8" coll="allgather" inplace="1" outofplace="0" minBytes="1048576" maxBytes="41943040">
-  <gpu id="0" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="1" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="2" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="3" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="4" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="5" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="6" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="7" recv="7" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="7" recv="7" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="7" recv="7" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="7" recv="7" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="7" recv="7" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="7" recv="7" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="7" recv="7" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="7" i_chunks="0" o_chunks="64" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="0" recv="0" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="0" recv="0" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="0" recv="0" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="0" recv="0" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="0" recv="0" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="0" recv="0" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="0" recv="0" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="1" recv="1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="1" recv="1" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="1" recv="1" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="1" recv="1" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="1" recv="1" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="1" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="1" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="16" dstbuf="o" dstoff="16" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="2" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="20" dstbuf="o" dstoff="20" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="2" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="17" dstbuf="o" dstoff="17" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="2" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="21" dstbuf="o" dstoff="21" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="2" recv="2" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="18" dstbuf="o" dstoff="18" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="2" recv="2" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="22" dstbuf="o" dstoff="22" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="2" recv="2" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="19" dstbuf="o" dstoff="19" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="2" recv="2" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="23" dstbuf="o" dstoff="23" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="24" dstbuf="o" dstoff="24" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="3" recv="3" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="28" dstbuf="o" dstoff="28" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="3" recv="3" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="25" dstbuf="o" dstoff="25" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="3" recv="3" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="29" dstbuf="o" dstoff="29" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="28" send="3" recv="3" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="26" dstbuf="o" dstoff="26" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="29" send="3" recv="3" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="30" dstbuf="o" dstoff="30" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="30" send="3" recv="3" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="27" dstbuf="o" dstoff="27" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="31" send="3" recv="3" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="32" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="32" dstbuf="o" dstoff="32" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="33" send="4" recv="4" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="36" dstbuf="o" dstoff="36" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="34" send="4" recv="4" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="33" dstbuf="o" dstoff="33" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="35" send="4" recv="4" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="37" dstbuf="o" dstoff="37" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="36" send="4" recv="4" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="34" dstbuf="o" dstoff="34" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="37" send="4" recv="4" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="38" dstbuf="o" dstoff="38" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="38" send="4" recv="4" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="35" dstbuf="o" dstoff="35" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="39" send="4" recv="4" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="39" dstbuf="o" dstoff="39" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="40" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="40" dstbuf="o" dstoff="40" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="41" send="5" recv="5" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="44" dstbuf="o" dstoff="44" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="42" send="5" recv="5" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="41" dstbuf="o" dstoff="41" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="43" send="5" recv="5" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="45" dstbuf="o" dstoff="45" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="44" send="5" recv="5" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="42" dstbuf="o" dstoff="42" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="45" send="5" recv="5" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="46" dstbuf="o" dstoff="46" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="46" send="5" recv="5" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="43" dstbuf="o" dstoff="43" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="47" send="5" recv="5" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="47" dstbuf="o" dstoff="47" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="48" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="56" dstbuf="o" dstoff="56" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="48" dstbuf="o" dstoff="48" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="49" send="6" recv="6" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="60" dstbuf="o" dstoff="60" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="52" dstbuf="o" dstoff="52" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="50" send="6" recv="6" chan="2">
-      <step s="0" type="s" srcbuf="o" srcoff="57" dstbuf="o" dstoff="57" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="49" dstbuf="o" dstoff="49" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="51" send="6" recv="6" chan="3">
-      <step s="0" type="s" srcbuf="o" srcoff="61" dstbuf="o" dstoff="61" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="53" dstbuf="o" dstoff="53" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="52" send="6" recv="6" chan="4">
-      <step s="0" type="s" srcbuf="o" srcoff="58" dstbuf="o" dstoff="58" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="50" dstbuf="o" dstoff="50" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="53" send="6" recv="6" chan="5">
-      <step s="0" type="s" srcbuf="o" srcoff="62" dstbuf="o" dstoff="62" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="54" dstbuf="o" dstoff="54" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="54" send="6" recv="6" chan="6">
-      <step s="0" type="s" srcbuf="o" srcoff="59" dstbuf="o" dstoff="59" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="51" dstbuf="o" dstoff="51" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="55" send="6" recv="6" chan="7">
-      <step s="0" type="s" srcbuf="o" srcoff="63" dstbuf="o" dstoff="63" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="55" dstbuf="o" dstoff="55" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-</algo>
diff --git a/tools/msccl-algorithms/allgather-8n-8kb-128kb.xml b/tools/msccl-algorithms/allgather-8n-8kb-128kb.xml
deleted file mode 100644
index 48afcb9b5b..0000000000
--- a/tools/msccl-algorithms/allgather-8n-8kb-128kb.xml
+++ /dev/null
@@ -1,692 +0,0 @@
-<!-- Copyright (c) Microsoft Corporation. -->
-<!-- Licensed under the MIT License. -->
-<algo name="all_gather_llm" proto="LL" nchannels="2" nchunksperloop="16" ngpus="8" coll="allgather" inplace="1" outofplace="0" minBytes="8193" maxBytes="131072">
-  <gpu id="0" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="1" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="2" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="3" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="4" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="5" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="6" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="7" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="7" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="7" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="7" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="7" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="0" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="-1" recv="0" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="-1" recv="1" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="-1" recv="1" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="-1" recv="2" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="3" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="-1" recv="3" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="-1" recv="4" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="-1" recv="4" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="-1" recv="5" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="-1" recv="5" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="-1" recv="6" chan="0">
-      <step s="0" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="-1" recv="6" chan="1">
-      <step s="0" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="0" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="15" send="0" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="16" send="1" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="17" send="1" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="18" send="2" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="19" send="2" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="20" send="3" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="21" send="3" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="22" send="4" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="23" send="4" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="24" send="5" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="25" send="5" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="26" send="6" recv="-1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="27" send="6" recv="-1" chan="1">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-</algo>
diff --git a/tools/msccl-algorithms/allgather-allpairs-16n-16tb.xml b/tools/msccl-algorithms/allgather-allpairs-16n-16tb.xml
deleted file mode 100644
index c92728a9e3..0000000000
--- a/tools/msccl-algorithms/allgather-allpairs-16n-16tb.xml
+++ /dev/null
@@ -1,994 +0,0 @@
-<algo name="allgather_allpairs" proto="LL" nchannels="1" nchunksperloop="16" ngpus="16" coll="allgather" inplace="1" outofplace="0" minBytes="1" maxBytes="327680">
-  <gpu id="0" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="1" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="2" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="3" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="4" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="5" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="6" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="7" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="8" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="9" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="10" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="11" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="12" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="13" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="14" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="15" recv="15" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-  <gpu id="15" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="0" recv="0" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="0" dstbuf="o" dstoff="0" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="1" send="1" recv="1" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="1" dstbuf="o" dstoff="1" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="2" send="2" recv="2" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="2" dstbuf="o" dstoff="2" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="3" send="3" recv="3" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="3" dstbuf="o" dstoff="3" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="4" send="4" recv="4" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="5" recv="5" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="5" dstbuf="o" dstoff="5" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="6" recv="6" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="6" dstbuf="o" dstoff="6" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="7" send="7" recv="7" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="7" dstbuf="o" dstoff="7" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="8" send="8" recv="8" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="8" dstbuf="o" dstoff="8" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="9" send="9" recv="9" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="9" dstbuf="o" dstoff="9" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="10" send="10" recv="10" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="11" send="11" recv="11" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="11" dstbuf="o" dstoff="11" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="12" send="12" recv="12" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="12" dstbuf="o" dstoff="12" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="13" send="13" recv="13" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="13" dstbuf="o" dstoff="13" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="14" send="14" recv="14" chan="0">
-      <step s="0" type="s" srcbuf="o" srcoff="15" dstbuf="o" dstoff="15" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-      <step s="1" type="r" srcbuf="o" srcoff="14" dstbuf="o" dstoff="14" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-</algo>