From 1250a5a259962514d31bb5f8148f1d0f0a82b946 Mon Sep 17 00:00:00 2001
From: Georgios Pinitas <georgios.pinitas@arm.com>
Date: Tue, 2 Jan 2018 13:27:37 +0000
Subject: COMPMID-767 : Propagate hints to subgraph.

-Propagates hints to subgraph.
-Fixes dispatching of apropriate optimized DepthwiseConvolution kernel
for OpenCL backend. NEON backend is altered to default to the generic
case until COMPMID-769 is addressed.

Change-Id: I544f05cd99a9ac253f1b19aa4e4bb222b8fdd087
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/114781
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
---
 src/graph/SubGraph.cpp                      |  4 ++--
 src/graph/nodes/BranchLayer.cpp             |  2 +-
 src/graph/operations/CLSimpleOperations.cpp |  4 ++--
 src/graph/operations/NESimpleOperations.cpp | 18 ++++--------------
 4 files changed, 9 insertions(+), 19 deletions(-)

(limited to 'src/graph')

diff --git a/src/graph/SubGraph.cpp b/src/graph/SubGraph.cpp
index 977cd4a4ae..f62b2617c5 100644
--- a/src/graph/SubGraph.cpp
+++ b/src/graph/SubGraph.cpp
@@ -52,13 +52,13 @@ void SubGraph::add_tensor_object(std::unique_ptr<ITensorObject> tensor)
     }
 }
 
-std::unique_ptr<Graph> SubGraph::construct(TargetHint hint, std::unique_ptr<ITensorObject> input, std::unique_ptr<ITensorObject> output)
+std::unique_ptr<Graph> SubGraph::construct(const GraphContext &ctx, std::unique_ptr<ITensorObject> input, std::unique_ptr<ITensorObject> output)
 {
     auto graph = arm_compute::support::cpp14::make_unique<Graph>();
 
     // Set hint
     // TODO(geopin01): store hints of sub-graph
-    graph->hints().set_target_hint(hint);
+    graph->hints() = ctx.hints();
 
     // Configure input
     if(_input == nullptr)
diff --git a/src/graph/nodes/BranchLayer.cpp b/src/graph/nodes/BranchLayer.cpp
index d062e4b791..eea0540741 100644
--- a/src/graph/nodes/BranchLayer.cpp
+++ b/src/graph/nodes/BranchLayer.cpp
@@ -151,7 +151,7 @@ std::unique_ptr<arm_compute::IFunction> BranchLayer::instantiate_node(GraphConte
         }
 
         // Construct sub_graph
-        auto g = sg->construct(ctx.hints().target_hint(), std::move(in), std::move(out));
+        auto g = sg->construct(ctx, std::move(in), std::move(out));
 
         // Register graph to function
         func->register_graph(std::move(g));
diff --git a/src/graph/operations/CLSimpleOperations.cpp b/src/graph/operations/CLSimpleOperations.cpp
index 8f2bf23ce3..61315e73b2 100644
--- a/src/graph/operations/CLSimpleOperations.cpp
+++ b/src/graph/operations/CLSimpleOperations.cpp
@@ -156,13 +156,13 @@ REGISTER_SIMPLE_OPERATION(CLDepthwiseConvolutionOperation, OPENCL, OperationType
     bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
     if(run_3x3_opt)
     {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer3x3>();
         depwthwise_conv->configure(in, weights, biases, out, conv_info);
         func = std::move(depwthwise_conv);
     }
     else
     {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer3x3>();
+        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();
         depwthwise_conv->configure(in, weights, biases, out, conv_info);
         func = std::move(depwthwise_conv);
     }
diff --git a/src/graph/operations/NESimpleOperations.cpp b/src/graph/operations/NESimpleOperations.cpp
index bb99e8da4b..88bf3ec0a0 100644
--- a/src/graph/operations/NESimpleOperations.cpp
+++ b/src/graph/operations/NESimpleOperations.cpp
@@ -149,23 +149,13 @@ REGISTER_SIMPLE_OPERATION(NEDepthwiseConvolutionOperation, NEON, OperationType::
     auto      *biases    = ctx.num_inputs() == 3 ? dynamic_cast<arm_compute::ITensor *>(ctx.input(2)) : nullptr;
     auto      *out       = dynamic_cast<arm_compute::ITensor *>(ctx.output(0));
     const auto conv_info = ctx.parameter<PadStrideInfo>("ConvolutionInfo");
-    const auto opt3x3    = ctx.parameter<bool>("Optimized3x3");
 
     // Create and configure function
     std::unique_ptr<arm_compute::IFunction> func;
-    bool                                    run_3x3_opt = opt3x3 && weights->info()->dimension(0) == 3;
-    if(run_3x3_opt)
-    {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
-        depwthwise_conv->configure(in, weights, biases, out, conv_info);
-        func = std::move(depwthwise_conv);
-    }
-    else
-    {
-        auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer3x3>();
-        depwthwise_conv->configure(in, weights, biases, out, conv_info);
-        func = std::move(depwthwise_conv);
-    }
+    // TODO (COMPMID-769): Add support for asymmetric padding in NEDepthwiseConvolutionLayer3x3 to enable opt3x3 support
+    auto depwthwise_conv = arm_compute::support::cpp14::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
+    depwthwise_conv->configure(in, weights, biases, out, conv_info);
+    func = std::move(depwthwise_conv);
 
     // Log info
     ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEDepthwiseConvolutionLayer"
-- 
cgit v1.2.1