Enable Transpose op for ACL neon (#7155)

This commit enables to support Transpose op for ACL neon.

Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>

diff --git a/runtimes/neurun/backend/acl_neon/KernelGenerator.cc b/runtimes/neurun/backend/acl_neon/KernelGenerator.cc
index f892ee9..3e4f09e 100644 (file)
--- a/runtimes/neurun/backend/acl_neon/KernelGenerator.cc
+++ b/runtimes/neurun/backend/acl_neon/KernelGenerator.cc
@@ -1119,8 +1119,53 @@ void KernelGenerator::visit(const model::operation::TransposeConvNode &node)
 
 void KernelGenerator::visit(const model::operation::TransposeNode &node)
 {
-  (void)node;
-  throw std::runtime_error("Not supported, yet");
+  const auto ofm_idx{node.getOutputs().at(0)};
+  const auto ifm_idx{node.getInputs().at(model::operation::TransposeNode::Input::INPUT)};
+  const auto perm{node.param().perm};
+
+  const auto rank = _ctx.at(ifm_idx).shape().rank();
+  std::vector<int32_t> pv;
+  const auto perm_base = _ctx.at(perm).data().base();
+  const int perm_size = _ctx.at(perm).shape().num_elements();
+
+  assert(perm_base != nullptr);
+  for (int32_t n = 0; n < perm_size; ++n)
+  {
+    const int32_t perm_value = *(reinterpret_cast<const int32_t *>(perm_base) + n);
+    assert(perm_value < rank);
+    pv.emplace_back(perm_value);
+  }
+
+  auto ofm_alloc = _tensor_builder->at(ofm_idx).get();
+  const auto ifm_alloc = _tensor_builder->at(ifm_idx).get();
+  const auto frontend_layout = _current_subg_layout;
+  const auto backend_layout = ifm_alloc->layout();
+
+  auto backend_pv = ::neurun::backend::acl_common::getARMComputePermutationVector(
+      rank, pv, frontend_layout, backend_layout);
+
+  std::unique_ptr<::arm_compute::IFunction> fn;
+
+  if (ifm_alloc->num_dimensions() <= 2 && ofm_alloc->num_dimensions() <= 2)
+  {
+    auto l = nnfw::cpp14::make_unique<::arm_compute::NETranspose>();
+
+    l->configure(ifm_alloc->handle(), ofm_alloc->handle());
+
+    fn = std::move(l);
+  }
+  else
+  {
+    auto l = nnfw::cpp14::make_unique<::arm_compute::NEPermute>();
+
+    l->configure(ifm_alloc->handle(), ofm_alloc->handle(), backend_pv);
+
+    fn = std::move(l);
+  }
+
+  auto acl_fn = asAclFunction(std::move(fn));
+
+  _execution_builder->append(std::move(acl_fn));
 }
 
 void KernelGenerator::visit(const model::operation::AddNode &node)

diff --git a/runtimes/neurun/backend/acl_neon/ShapeFixer.cc b/runtimes/neurun/backend/acl_neon/ShapeFixer.cc
index 046a18a..1ef911d 100644 (file)
--- a/runtimes/neurun/backend/acl_neon/ShapeFixer.cc
+++ b/runtimes/neurun/backend/acl_neon/ShapeFixer.cc
@@ -206,6 +206,8 @@ void ShapeFixer::visit(const model::operation::SubNode &node)
 
 void ShapeFixer::visit(const model::operation::TransposeConvNode &) { /* DO NOTHING */}
 
+void ShapeFixer::visit(const model::operation::TransposeNode &) { /* DO NOTHING */}
+
 void ShapeFixer::visit(const model::operation::AddNode &node)
 {
   const auto lhs_index{node.getInputs().at(model::operation::AddNode::Input::LHS)};

diff --git a/runtimes/neurun/backend/acl_neon/ShapeFixer.h b/runtimes/neurun/backend/acl_neon/ShapeFixer.h
index 2c3987d..ee0f4f7 100644 (file)
--- a/runtimes/neurun/backend/acl_neon/ShapeFixer.h
+++ b/runtimes/neurun/backend/acl_neon/ShapeFixer.h
@@ -65,6 +65,7 @@ public:
   void visit(const model::operation::SubNode &) override;
   void visit(const model::operation::StridedSliceNode &) override;
   void visit(const model::operation::TransposeConvNode &) override;
+  void visit(const model::operation::TransposeNode &) override;
   void visit(const model::operation::AddNode &) override;
   void visit(const model::operation::DivNode &) override;
   void visit(const model::operation::ComparisonNode &) override;

diff --git a/tests/nnapi/nnapi_gtest.skip.armv7l-linux.acl_neon b/tests/nnapi/nnapi_gtest.skip.armv7l-linux.acl_neon
index 97792df..f94f078 100644 (file)
--- a/tests/nnapi/nnapi_gtest.skip.armv7l-linux.acl_neon
+++ b/tests/nnapi/nnapi_gtest.skip.armv7l-linux.acl_neon
@@ -28,7 +28,6 @@ GeneratedTests.svdf*
 GeneratedTests.tanh_
 GeneratedTests.batch_to_space*
 GeneratedTests.space_to_batch*
-GeneratedTests.transpose*
 GeneratedTests.cast_ex*
 GeneratedTests.gather_ex*
 GeneratedTests.strided_slice_ex*

diff --git a/tests/scripts/neurun_frameworktest_list.armv7l.acl_neon.txt b/tests/scripts/neurun_frameworktest_list.armv7l.acl_neon.txt
index d5c1149..6b4e210 100644 (file)
--- a/tests/scripts/neurun_frameworktest_list.armv7l.acl_neon.txt
+++ b/tests/scripts/neurun_frameworktest_list.armv7l.acl_neon.txt
@@ -14,6 +14,7 @@ reshape
 softmax
 sqrt
 tanh
+transpose
 transpose_conv
 MODELS/inception_module
 MODELS/mobilenet