Merge branch 'develop' into jit-vector-reduce

examples/vision/python_yolov4/yolov4_inference.ipynb

@@ -24,16 +24,16 @@
     "import os.path\n",
     "\n",
     "if not os.path.exists(\"./utilities/coco.names\"):\n",
-    "    !wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/yolov4/dependencies/coco.names -P ./utilities/\n",
+    "    !wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/yolov4/dependencies/coco.names -P ./utilities/\n",
     "if not os.path.exists(\"./utilities/yolov4_anchors.txt\"):\n",
-    "    !wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/yolov4/dependencies/yolov4_anchors.txt -P ./utilities/\n",
+    "    !wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/yolov4/dependencies/yolov4_anchors.txt -P ./utilities/\n",
     "if not os.path.exists(\"./utilities/input.jpg\"):\n",
     "    # The image used is from the COCO dataset (https://cocodataset.org/#explore)\n",
     "    # Other images can be tested by replacing the link below\n",
     "    image_link = \"https://farm3.staticflickr.com/2009/2306189268_88cc86b30f_z.jpg\"\n",
     "    !wget -O ./utilities/input.jpg $image_link\n",
     "if not os.path.exists(\"./utilities/yolov4.onnx\"):\n",
-    "    !wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/yolov4/model/yolov4.onnx -P ./utilities/"
+    "    !wget https://github.com/onnx/models/raw/main/vision/object_detection_segmentation/yolov4/model/yolov4.onnx -P ./utilities/"
    ]
   },
   {

src/propagate_constant.cpp

@@ -3,6 +3,7 @@
 #include <migraphx/matcher.hpp>
 #include <migraphx/literal.hpp>
 #include <migraphx/functional.hpp>
+#include <migraphx/par_for.hpp>
 #include <unordered_set>
 
 namespace migraphx {
@@ -20,33 +21,42 @@ bool skip_propogate(instruction_ref ins)
     return false;
 }
 
+bool is_const(instruction_ref ins) { return ins->can_eval() and not skip_propogate(ins); }
+
 void propagate_constant::apply(module& m) const
 {
+    std::unordered_set<instruction_ref> const_instrs;
+    auto last = std::prev(m.end());
+
+    // Find instructions that can be evaluated to a literal
     for(auto i : iterator_for(m))
     {
-        if(i->name() != "@literal")
-            continue;
-        if(i->outputs().empty())
-            continue;
-        fix([&](auto self, auto ins) {
-            std::unordered_set<instruction_ref> children(ins->outputs().begin(),
-                                                         ins->outputs().end());
-            for(auto child : children)
-            {
-                if(child->name() == "@literal" or skip_propogate(child))
-                {
-                    self(child);
+        if(is_const(i) and i != last)
             continue;
+
+        std::copy_if(
+            i->inputs().begin(),
+            i->inputs().end(),
+            std::inserter(const_instrs, const_instrs.begin()),
+            [&](const instruction_ref ins) { return is_const(ins) and ins->name() != "@literal"; });
     }
-                auto r = child->eval();
-                if(not r.empty())
+
+    // Compute literals in parallel
+    std::vector<instruction_ref> const_instrs_vec{const_instrs.begin(), const_instrs.end()};
+    std::vector<argument> literals(const_instrs_vec.size());
+    par_for(const_instrs_vec.size(), 1, [&](const auto i) {
+        literals[i] = const_instrs_vec[i]->eval();
+    });
+
+    // Replace instructions in m
+    for(size_t i = 0; i < const_instrs_vec.size(); i++)
     {
-                    assert(r.get_shape() == child->get_shape());
-                    auto l = m.add_literal(r.get_shape(), r.data());
-                    self(m.replace_instruction(child, l));
-                }
+        if(not literals[i].empty())
+        {
+            assert(literals[i].get_shape() == const_instrs_vec[i]->get_shape());
+            auto l = m.add_literal(literals[i].get_shape(), literals[i].data());
+            m.replace_instruction(const_instrs_vec[i], l);
         }
-        })(i);
     }
 }