From 50f22e8317c95f989ec0b27f7842ce174c8aea97 Mon Sep 17 00:00:00 2001
From: Pedram Alizadeh <pmohamma@amd.com>
Date: Thu, 14 Mar 2024 12:18:26 -0400
Subject: [PATCH] msccl algorithms tuning for allgather on MI300 (#1110)

---
 ...er-8n-0-1kb.xml => allgather-8n-0-8kb.xml} |    2 +-
 ...1mb-4mb.xml => allgather-8n-128kb-1mb.xml} |    2 +-
 ...4mb-16mb.xml => allgather-8n-1mb-40mb.xml} |    2 +-
 .../msccl-algorithms/allgather-8n-4kb-8kb.xml |  692 ---------
 ...1kb-4kb.xml => allgather-8n-8kb-128kb.xml} |    2 +-
 .../msccl-algorithms/allgather-8n-8kb-1mb.xml | 1364 -----------------
 6 files changed, 4 insertions(+), 2060 deletions(-)
 rename tools/msccl-algorithms/{allgather-8n-0-1kb.xml => allgather-8n-0-8kb.xml} (99%)
 rename tools/msccl-algorithms/{allgather-8n-1mb-4mb.xml => allgather-8n-128kb-1mb.xml} (99%)
 rename tools/msccl-algorithms/{allgather-8n-4mb-16mb.xml => allgather-8n-1mb-40mb.xml} (99%)
 delete mode 100644 tools/msccl-algorithms/allgather-8n-4kb-8kb.xml
 rename tools/msccl-algorithms/{allgather-8n-1kb-4kb.xml => allgather-8n-8kb-128kb.xml} (99%)
 delete mode 100644 tools/msccl-algorithms/allgather-8n-8kb-1mb.xml

diff --git a/tools/msccl-algorithms/allgather-8n-0-1kb.xml b/tools/msccl-algorithms/allgather-8n-0-8kb.xml
similarity index 99%
rename from tools/msccl-algorithms/allgather-8n-0-1kb.xml
rename to tools/msccl-algorithms/allgather-8n-0-8kb.xml
index 8588571562..6b43e13b4d 100644
--- a/tools/msccl-algorithms/allgather-8n-0-1kb.xml
+++ b/tools/msccl-algorithms/allgather-8n-0-8kb.xml
@@ -1,6 +1,6 @@
 <!-- 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="1023">
+<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"/>
diff --git a/tools/msccl-algorithms/allgather-8n-1mb-4mb.xml b/tools/msccl-algorithms/allgather-8n-128kb-1mb.xml
similarity index 99%
rename from tools/msccl-algorithms/allgather-8n-1mb-4mb.xml
rename to tools/msccl-algorithms/allgather-8n-128kb-1mb.xml
index 2da0b6cd9f..03e14f2feb 100644
--- a/tools/msccl-algorithms/allgather-8n-1mb-4mb.xml
+++ b/tools/msccl-algorithms/allgather-8n-128kb-1mb.xml
@@ -1,6 +1,6 @@
 <!-- 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="1048576" maxBytes="4194303">
+<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"/>
diff --git a/tools/msccl-algorithms/allgather-8n-4mb-16mb.xml b/tools/msccl-algorithms/allgather-8n-1mb-40mb.xml
similarity index 99%
rename from tools/msccl-algorithms/allgather-8n-4mb-16mb.xml
rename to tools/msccl-algorithms/allgather-8n-1mb-40mb.xml
index 8a6f4939ad..da1de14d6c 100644
--- a/tools/msccl-algorithms/allgather-8n-4mb-16mb.xml
+++ b/tools/msccl-algorithms/allgather-8n-1mb-40mb.xml
@@ -1,6 +1,6 @@
 <!-- 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="4194303" maxBytes="16777216">
+<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"/>
diff --git a/tools/msccl-algorithms/allgather-8n-4kb-8kb.xml b/tools/msccl-algorithms/allgather-8n-4kb-8kb.xml
deleted file mode 100644
index c231dbe967..0000000000
--- a/tools/msccl-algorithms/allgather-8n-4kb-8kb.xml
+++ /dev/null
@@ -1,692 +0,0 @@
-<!-- Copyright (c) Microsoft Corporation. -->
-<!-- Licensed under the MIT License. -->
-<algo name="all_gather_llm" proto="LL" nchannels="16" nchunksperloop="16" ngpus="8" coll="allgather" inplace="1" outofplace="0" minBytes="4096" maxBytes="8191">
-  <gpu id="0" i_chunks="0" o_chunks="16" s_chunks="0">
-    <tb id="0" send="-1" recv="1" 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="1" send="-1" recv="1" chan="9">
-      <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="2">
-      <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="10">
-      <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="3">
-      <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="11">
-      <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="4">
-      <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="12">
-      <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="5">
-      <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="13">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="8">
-      <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="8">
-      <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="8">
-      <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="8">
-      <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="8">
-      <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="8">
-      <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="8">
-      <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="8">
-      <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="2">
-      <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="10">
-      <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="3">
-      <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="11">
-      <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="4">
-      <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="12">
-      <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="5">
-      <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="13">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="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="15" send="0" recv="-1" chan="9">
-      <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="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="17" send="2" recv="-1" chan="9">
-      <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="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="19" send="3" recv="-1" chan="9">
-      <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="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="21" send="4" recv="-1" chan="9">
-      <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="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="23" send="5" recv="-1" chan="9">
-      <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="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="25" send="6" recv="-1" chan="9">
-      <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="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="27" send="7" recv="-1" chan="9">
-      <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="8">
-      <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="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="3" send="-1" recv="1" chan="9">
-      <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="3">
-      <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="11">
-      <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="4">
-      <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="12">
-      <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="5">
-      <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="13">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="2">
-      <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="10">
-      <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="2">
-      <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="10">
-      <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="2">
-      <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="10">
-      <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="2">
-      <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="10">
-      <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="2">
-      <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="10">
-      <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="2">
-      <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="10">
-      <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="2">
-      <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="10">
-      <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="8">
-      <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="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="3" send="-1" recv="1" chan="9">
-      <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="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="10">
-      <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="4">
-      <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="12">
-      <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="5">
-      <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="13">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="3">
-      <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="11">
-      <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="3">
-      <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="11">
-      <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="3">
-      <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="11">
-      <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="3">
-      <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="11">
-      <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="3">
-      <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="11">
-      <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="3">
-      <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="11">
-      <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="3">
-      <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="11">
-      <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="8">
-      <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="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="3" send="-1" recv="1" chan="9">
-      <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="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="10">
-      <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="3">
-      <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="11">
-      <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="5">
-      <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="13">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="4">
-      <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="12">
-      <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="4">
-      <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="12">
-      <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="4">
-      <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="12">
-      <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="4">
-      <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="12">
-      <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="4">
-      <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="12">
-      <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="4">
-      <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="12">
-      <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="4">
-      <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="12">
-      <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="8">
-      <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="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="3" send="-1" recv="1" chan="9">
-      <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="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="10">
-      <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="3">
-      <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="11">
-      <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="4">
-      <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="12">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="5">
-      <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="13">
-      <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="5">
-      <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="13">
-      <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="5">
-      <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="13">
-      <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="5">
-      <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="13">
-      <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="5">
-      <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="13">
-      <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="5">
-      <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="13">
-      <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="5">
-      <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="13">
-      <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="8">
-      <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="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="3" send="-1" recv="1" chan="9">
-      <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="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="10">
-      <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="3">
-      <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="11">
-      <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="4">
-      <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="12">
-      <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="5">
-      <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="13">
-      <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="7">
-      <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="15">
-      <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="6">
-      <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="14">
-      <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="6">
-      <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="14">
-      <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="6">
-      <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="14">
-      <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="6">
-      <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="14">
-      <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="6">
-      <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="14">
-      <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="6">
-      <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="14">
-      <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="6">
-      <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="14">
-      <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="8">
-      <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="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="3" send="-1" recv="1" chan="9">
-      <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="4" dstbuf="o" dstoff="4" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="5" send="-1" recv="2" chan="10">
-      <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="3">
-      <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="11">
-      <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="4">
-      <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="12">
-      <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="5">
-      <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="13">
-      <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="6">
-      <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="14">
-      <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="7">
-      <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="15">
-      <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="7">
-      <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="15">
-      <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="7">
-      <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="15">
-      <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="7">
-      <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="15">
-      <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="7">
-      <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="15">
-      <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="7">
-      <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="15">
-      <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="7">
-      <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="15">
-      <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-8n-1kb-4kb.xml b/tools/msccl-algorithms/allgather-8n-8kb-128kb.xml
similarity index 99%
rename from tools/msccl-algorithms/allgather-8n-1kb-4kb.xml
rename to tools/msccl-algorithms/allgather-8n-8kb-128kb.xml
index a5fb9bcbf6..48afcb9b5b 100644
--- a/tools/msccl-algorithms/allgather-8n-1kb-4kb.xml
+++ b/tools/msccl-algorithms/allgather-8n-8kb-128kb.xml
@@ -1,6 +1,6 @@
 <!-- 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="1024" maxBytes="4095">
+<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"/>
diff --git a/tools/msccl-algorithms/allgather-8n-8kb-1mb.xml b/tools/msccl-algorithms/allgather-8n-8kb-1mb.xml
deleted file mode 100644
index 8f1f7825ce..0000000000
--- a/tools/msccl-algorithms/allgather-8n-8kb-1mb.xml
+++ /dev/null
@@ -1,1364 +0,0 @@
-<!-- Copyright (c) Microsoft Corporation. -->
-<!-- Licensed under the MIT License. -->
-<algo name="all_gather_llm" proto="LL" nchannels="4" nchunksperloop="32" ngpus="8" coll="allgather" inplace="1" outofplace="0" minBytes="8192" maxBytes="1048575">
-  <gpu id="0" i_chunks="0" o_chunks="32" s_chunks="0">
-    <tb id="0" send="-1" recv="1" 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="1" send="-1" recv="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="5" send="-1" recv="2" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="9" send="-1" recv="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="8" dstbuf="o" dstoff="8" 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="10" dstbuf="o" dstoff="10" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-    <tb id="6" send="-1" recv="2" chan="2">
-      <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="3">
-      <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="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="9" send="-1" recv="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="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="29" send="0" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="33" send="2" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="37" send="3" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="41" send="4" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="45" send="5" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="49" send="6" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="53" send="7" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="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="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="9" send="-1" recv="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="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="29" send="0" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="33" send="1" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="37" send="3" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="41" send="4" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="45" send="5" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="49" send="6" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="53" send="7" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="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="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="2" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="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="29" send="0" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="33" send="1" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="37" send="2" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="41" send="4" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="45" send="5" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="49" send="6" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="53" send="7" recv="-1" chan="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="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="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="2" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="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="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="2" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="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="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="2" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="2">
-      <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="3">
-      <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="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="1" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="2" chan="1">
-      <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="2">
-      <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="3">
-      <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="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="3" chan="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <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="0">
-      <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="1">
-      <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="2">
-      <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="3">
-      <step s="0" type="s" srcbuf="o" srcoff="31" dstbuf="o" dstoff="31" cnt="1" depid="-1" deps="-1" hasdep="0"/>
-    </tb>
-  </gpu>
-</algo>