From 2ba1235cad86432f61deb3bddd46b40e7b163c31 Mon Sep 17 00:00:00 2001
From: =?utf8?q?=EC=9D=B4=ED=95=9C=EC=A2=85/=EB=8F=99=EC=9E=91=EC=A0=9C?=
 =?utf8?q?=EC=96=B4Lab=28SR=29/Engineer/=EC=82=BC=EC=84=B1=EC=A0=84?=
 =?utf8?q?=EC=9E=90?= <hanjoung.lee@samsung.com>
Date: Thu, 19 Jul 2018 15:38:57 +0900
Subject: [PATCH] [newrt] Support Concat from acl_cl backend (#2010)

Enable Concat from acl_cl backend. So now we can run inception v3
with acl_cl backend only.

Signed-off-by: Hanjoung Lee <hanjoung.lee@samsung.com>
---
 .../src/internal/arm_compute/StageGenerator.cc     | 41 +++++++++++++++++++---
 .../src/internal/kernel/acl_cl/ConcatLayer.cc      |  2 +-
 .../test_driver/newruntime_frameworktest_list.txt  |  1 +
 3 files changed, 39 insertions(+), 5 deletions(-)

diff --git a/runtimes/neurun/src/internal/arm_compute/StageGenerator.cc b/runtimes/neurun/src/internal/arm_compute/StageGenerator.cc
index 333a2bd..ca4c858 100644
--- a/runtimes/neurun/src/internal/arm_compute/StageGenerator.cc
+++ b/runtimes/neurun/src/internal/arm_compute/StageGenerator.cc
@@ -7,6 +7,8 @@
 #include <arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h>
 #include <arm_compute/runtime/CL/functions/CLSoftmaxLayer.h>
 
+#include "internal/kernel/acl_cl/ConcatLayer.h"
+
 #include "internal/Padding.h"
 #include "internal/Model.h"
 
@@ -355,11 +357,42 @@ Stage StageGenerator::generate(const ::internal::tflite::op::AvgPool2D::implicit
 
 Stage StageGenerator::generate(const ::internal::tflite::op::Concat::Node &node)
 {
-  throw std::runtime_error{"NYI - StageGenerator::generate for 'Concat'"};
+  const ::internal::tflite::operand::Index ofm_index{node.param().ofm_index};
+  const ::internal::tflite::operand::Index axis_index{node.param().axis_index};
+
+  struct Param
+  {
+    int32_t output_index;
+    std::vector<int32_t> input_indexes;
+
+    int32_t axis;
+  };
+
+  Param param;
+
+  param.output_index = node.param().ofm_index;
+  param.input_indexes = node.param().ifm_indexes;
+  param.axis = _ctx.at(axis_index).asScalar<int32_t>();
+
+  VERBOSE(Concat) << param.axis << std::endl;
 
-  return [](IExecutionBuilder &builder) {
-    // NOTE arm_compute does not have Concat operation
-    // TODO Implement
+  auto tensors = _tensor_builder;
+
+  return [tensors, param](IExecutionBuilder &builder) {
+    auto output_alloc = tensors->at(::internal::tflite::operand::Index{param.output_index}).get();
+
+    std::vector<::arm_compute::ICLTensor *> input_allocs;
+    for (auto ifm_ind : param.input_indexes)
+    {
+      input_allocs.emplace_back(tensors->at(::internal::tflite::operand::Index{ifm_ind}).get());
+    }
+
+    std::unique_ptr<::internal::kernel::acl_cl::ConcatLayer> fn{
+        new ::internal::kernel::acl_cl::ConcatLayer};
+
+    fn->configure(input_allocs, param.axis, output_alloc);
+
+    builder.append(std::move(fn));
   };
 }
 
diff --git a/runtimes/neurun/src/internal/kernel/acl_cl/ConcatLayer.cc b/runtimes/neurun/src/internal/kernel/acl_cl/ConcatLayer.cc
index 2b98059..55649cc 100644
--- a/runtimes/neurun/src/internal/kernel/acl_cl/ConcatLayer.cc
+++ b/runtimes/neurun/src/internal/kernel/acl_cl/ConcatLayer.cc
@@ -34,7 +34,7 @@ bool matchSizeExceptAxis(const ::arm_compute::ICLTensor *t1, const ::arm_compute
   {
     if (axis == i)
       continue;
-    if (t1->info()->dimension(i) == t2->info()->dimension(i))
+    if (t1->info()->dimension(i) != t2->info()->dimension(i))
       return false;
   }
   return true;
diff --git a/tools/test_driver/newruntime_frameworktest_list.txt b/tools/test_driver/newruntime_frameworktest_list.txt
index 80c8173..a5f612b 100644
--- a/tools/test_driver/newruntime_frameworktest_list.txt
+++ b/tools/test_driver/newruntime_frameworktest_list.txt
@@ -5,3 +5,4 @@ convolution2
 maxpool1
 maxpool2
 softmax
+inception
-- 
2.7.4