commit 59631a174e1b5ef23bd3a0102f60b57c99502766
author Giorgio Arena <giorgio.arena@arm.com> Wed May 02 13:59:04 2018 +0100
committer Anthony Barbier <anthony.barbier@arm.com> Fri Nov 02 16:51:50 2018 +0000
tree 5d8e15d7a3b65e5071db82e2937ee1808953823f
parent ef9e05978ab008e533cc76a8e6f10c9e86a880c1

COMPMID-1104 Add fast math hint in the graph API

Change-Id: I83db135fa94c6884e080f0229a9b6430d908c029
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/129823
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

src/graph/GraphBuilder.cpp

diff --git a/src/graph/GraphBuilder.cpp b/src/graph/GraphBuilder.cpp
index 56b31c7..df94d0b 100644
--- a/src/graph/GraphBuilder.cpp
+++ b/src/graph/GraphBuilder.cpp
@@ -81,7 +81,7 @@
 }
 
 NodeID create_grouped_convolution(Graph &g, NodeParams &params, NodeIdxPair input, NodeID weights, NodeID bias,
-                                  PadStrideInfo conv_info, ConvolutionMethod method, unsigned int num_groups)
+                                  PadStrideInfo conv_info, ConvolutionMethod method, FastMathHint fast_math_hint, unsigned int num_groups)
 {
     bool has_bias = (bias != EmptyNodeID);
 
@@ -102,7 +102,7 @@
     std::vector<NodeIdxPair> convolution_outputs;
     for(unsigned int i = 0; i < num_groups; ++i)
     {
-        NodeID conv_nid = g.add_node<ConvolutionLayerNode>(conv_info, method);
+        NodeID conv_nid = g.add_node<ConvolutionLayerNode>(conv_info, method, fast_math_hint);
         g.add_connection(input_split, i, conv_nid, 0);
         g.add_connection(weights_split, i, conv_nid, 1);
         if(has_bias)
@@ -205,7 +205,7 @@
 
 NodeID GraphBuilder::add_convolution_node(Graph &g, NodeParams params, NodeIdxPair input,
                                           Size2D kernel_spatial_extend, unsigned int depth, PadStrideInfo conv_info,
-                                          unsigned int num_groups, ConvolutionMethod method,
+                                          unsigned int num_groups, ConvolutionMethod method, FastMathHint fast_math_hint,
                                           ITensorAccessorUPtr weights_accessor, ITensorAccessorUPtr bias_accessor,
                                           const QuantizationInfo weights_quant_info,
                                           const QuantizationInfo out_quant_info)
@@ -245,7 +245,7 @@
     if(num_groups == 1)
     {
         // Create convolution node and connect
-        NodeID conv_nid = g.add_node<ConvolutionLayerNode>(conv_info, method, out_quant_info);
+        NodeID conv_nid = g.add_node<ConvolutionLayerNode>(conv_info, method, fast_math_hint, out_quant_info);
         g.add_connection(input.node_id, input.index, conv_nid, 0);
         g.add_connection(w_nid, 0, conv_nid, 1);
         if(has_bias)
@@ -258,7 +258,7 @@
     }
     else
     {
-        return create_grouped_convolution(g, params, input, w_nid, b_nid, conv_info, method, num_groups);
+        return create_grouped_convolution(g, params, input, w_nid, b_nid, conv_info, method, fast_math_hint, num_groups);
     }
 }
 

src/graph/backends/CL/CLFunctionsFactory.cpp

diff --git a/src/graph/backends/CL/CLFunctionsFactory.cpp b/src/graph/backends/CL/CLFunctionsFactory.cpp
index ece6364..4626cb5 100644
--- a/src/graph/backends/CL/CLFunctionsFactory.cpp
+++ b/src/graph/backends/CL/CLFunctionsFactory.cpp
@@ -166,6 +166,7 @@
 
     const PadStrideInfo     conv_info      = node.convolution_info();
     const ConvolutionMethod conv_algorithm = node.convolution_method();
+    const bool              fast_math      = node.fast_math_hint() == FastMathHint::ENABLED;
 
     // Create and configure function (we assume that functions have been validated before creation)
     std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, Target::CL);
@@ -175,7 +176,7 @@
     if(conv_algorithm == ConvolutionMethod::WINOGRAD)
     {
         std::tie(func, func_name) = create_named_memory_managed_function<CLWinogradConvolutionLayer>(
-                                        std::string("CLWinogradConvolutionLayer"), mm, input, weights, biases, output, conv_info);
+                                        std::string("CLWinogradConvolutionLayer"), mm, input, weights, biases, output, conv_info, ActivationLayerInfo(), fast_math);
     }
     else if(conv_algorithm == ConvolutionMethod::DIRECT)
     {
@@ -190,7 +191,7 @@
     else
     {
         std::tie(func, func_name) = create_named_memory_managed_function<CLConvolutionLayer>(std::string("CLConvolutionLayer"), mm,
-                                                                                             input, weights, biases, output, conv_info);
+                                                                                             input, weights, biases, output, conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), fast_math);
     }
 
     // Log info

src/graph/nodes/ConvolutionLayerNode.cpp

diff --git a/src/graph/nodes/ConvolutionLayerNode.cpp b/src/graph/nodes/ConvolutionLayerNode.cpp
index eaf1f7f..6c31a6b 100644
--- a/src/graph/nodes/ConvolutionLayerNode.cpp
+++ b/src/graph/nodes/ConvolutionLayerNode.cpp
@@ -32,8 +32,8 @@
 {
 namespace graph
 {
-ConvolutionLayerNode::ConvolutionLayerNode(PadStrideInfo info, ConvolutionMethod method, QuantizationInfo out_quant_info)
-    : _info(std::move(info)), _method(method), _out_quant_info(out_quant_info)
+ConvolutionLayerNode::ConvolutionLayerNode(PadStrideInfo info, ConvolutionMethod method, FastMathHint fast_math_hint, QuantizationInfo out_quant_info)
+    : _info(std::move(info)), _method(method), _fast_math_hint(fast_math_hint), _out_quant_info(out_quant_info)
 {
     _input_edges.resize(3, EmptyEdgeID);
     _outputs.resize(1, NullTensorID);
@@ -49,6 +49,16 @@
     return _method;
 }
 
+void ConvolutionLayerNode::set_fast_math_hint(FastMathHint hint)
+{
+    _fast_math_hint = hint;
+}
+
+FastMathHint ConvolutionLayerNode::fast_math_hint() const
+{
+    return _fast_math_hint;
+}
+
 PadStrideInfo ConvolutionLayerNode::convolution_info() const
 {
     return _info;