Merge branch 'master' into cyphers/opsdesc

src/contrib/mlir/lowerer.cpp

@@ -43,6 +43,8 @@ namespace
     using namespace mlir::edsc::op;
     using namespace ngraph::runtime;
     using namespace ngraph::runtime::ngmlir;
+    // Index notation to generate standard (i.e., non-affine) loads and stores.
+    using StdIndexedValue = TemplatedIndexedValue<intrinsics::std_load, intrinsics::std_store>;
 
     class DialectLoweringPass;
 
@@ -682,7 +684,8 @@ namespace
         // Create view to write into result.
         MemRefView vRes(result), vParams(params), vIndices(indices);
         // Indexed Values
-        IndexedValue iRes(result), iParams(params), iIndices(indices);
+        IndexedValue iRes(result), iIndices(indices);
+        StdIndexedValue iParams(params);
 
         // Construct outer loop for params dims. Exclude the axis dim.
         SmallVector<ValueHandle, 4> paramsLbs, paramsUbs;
@@ -894,7 +897,8 @@ namespace
         // Views
         MemRefView vRes(result), vArg(arg);
         // Index Values
-        IndexedValue iRes(result), iArg(arg);
+        StdIndexedValue iRes(result), stdArg(arg);
+        IndexedValue affineArg(arg);
         // Bounds Index Handles
         auto resLbs = vRes.getLbs();
         auto resUbs = vRes.getUbs();
@@ -944,9 +948,9 @@ namespace
                 ValueHandle newRedIdx =
                     std::is_same<RedOp, NGArgMinRedOp>()
                         ? edsc::intrinsics::select(
-                              iArg(allIVs) < iArg(tempIVs), allIVs[axis], currRedIdx)
+                              affineArg(allIVs) < stdArg(tempIVs), allIVs[axis], currRedIdx)
                         : edsc::intrinsics::select(
-                              iArg(tempIVs) < iArg(allIVs), allIVs[axis], currRedIdx);
+                              stdArg(tempIVs) < affineArg(allIVs), allIVs[axis], currRedIdx);
 
                 iRes(nonRedIVs) = ValueHandle::create<IndexCastOp>(newRedIdx, resTy);
             });

src/ngraph/pass/constant_folding.hpp

@@ -49,7 +49,9 @@ public:
         SLICE,
         DYN_SLICE,
         DYN_RESHAPE,
-        TRANSPOSE
+        TRANSPOSE,
+        RANGE,
+        SELECT
     };
 
     ConstantFolding(const ngraph::BuildNodeExecutorMap& cfmap = ngraph::BuildNodeExecutorMap())
@@ -74,6 +76,8 @@ public:
         construct_constant_dyn_slice();
         construct_constant_dyn_reshape();
         construct_constant_transpose();
+        construct_constant_range();
+        construct_constant_select();
     }
 
     //this allows to specify the order in which matchers will be run
@@ -105,6 +109,8 @@ public:
             case CFTransformations::DYN_SLICE: construct_constant_dyn_slice(); break;
             case CFTransformations::DYN_RESHAPE: construct_constant_dyn_reshape(); break;
             case CFTransformations::TRANSPOSE: construct_constant_transpose(); break;
+            case CFTransformations::RANGE: construct_constant_range(); break;
+            case CFTransformations::SELECT: construct_constant_select(); break;
             }
         }
     }
@@ -128,6 +134,8 @@ private:
     void construct_constant_dyn_slice();
     void construct_constant_dyn_reshape();
     void construct_constant_transpose();
+    void construct_constant_range();
+    void construct_constant_select();
 
     ngraph::BuildNodeExecutorMap m_cfmap;
 };

src/ngraph/pass/dyn_elimination.cpp

@@ -30,6 +30,7 @@
 #include "ngraph/op/slice.hpp"
 #include "ngraph/pattern/matcher.hpp"
 #include "ngraph/pattern/op/label.hpp"
+#include "ngraph/runtime/reference/range.hpp"
 #include "ngraph/slice_plan.hpp"
 
 using namespace std;
@@ -342,11 +343,10 @@ void pass::DynElimination::construct_dyn_reshape()
 }
 
 template <typename T>
-std::shared_ptr<op::Constant>
-    make_range_replacement_integral(const element::Type& et,
-                                    const Shape& shape,
-                                    const std::shared_ptr<op::Constant>& start_arg,
-                                    const std::shared_ptr<op::Constant>& step_arg)
+std::shared_ptr<op::Constant> make_range_replacement(const element::Type& et,
+                                                     const Shape& shape,
+                                                     const std::shared_ptr<op::Constant>& start_arg,
+                                                     const std::shared_ptr<op::Constant>& step_arg)
 {
     std::vector<T> elements(shape_size(shape));
     std::vector<T> start_vec = start_arg->get_vector<T>();
@@ -354,40 +354,7 @@ std::shared_ptr<op::Constant>
 
     NGRAPH_CHECK(start_vec.size() == 1 && step_vec.size() == 1);
 
-    T start = start_vec[0];
-    T step = step_vec[0];
-
-    T val = start;
-
-    for (size_t i = 0; i < elements.size(); i++)
-    {
-        elements[i] = val;
-        val = val + step;
-    }
-
-    return make_shared<op::Constant>(et, shape, elements);
-}
-
-template <typename T>
-std::shared_ptr<op::Constant>
-    make_range_replacement_floating(const element::Type& et,
-                                    const Shape& shape,
-                                    const std::shared_ptr<op::Constant>& start_arg,
-                                    const std::shared_ptr<op::Constant>& step_arg)
-{
-    std::vector<T> elements(shape_size(shape));
-    std::vector<T> start_vec = start_arg->get_vector<T>();
-    std::vector<T> step_vec = step_arg->get_vector<T>();
-
-    NGRAPH_CHECK(start_vec.size() == 1 && step_vec.size() == 1);
-
-    T start = start_vec[0];
-    T step = step_vec[0];
-
-    for (size_t i = 0; i < elements.size(); i++)
-    {
-        elements[i] = start + (static_cast<T>(i) * step);
-    }
+    runtime::reference::range<T>(start_vec.data(), step_vec.data(), shape, elements.data());
 
     return make_shared<op::Constant>(et, shape, elements);
 }
@@ -426,40 +393,40 @@ void pass::DynElimination::construct_range()
         switch (et.get_type_enum())
         {
         case element::Type_t::bf16:
-            replacement = make_range_replacement_floating<bfloat16>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<bfloat16>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::f16:
-            replacement = make_range_replacement_floating<float16>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<float16>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::f32:
-            replacement = make_range_replacement_floating<float>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<float>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::f64:
-            replacement = make_range_replacement_floating<double>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<double>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::i8:
-            replacement = make_range_replacement_integral<int8_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<int8_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::i16:
-            replacement = make_range_replacement_integral<int16_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<int16_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::i32:
-            replacement = make_range_replacement_integral<int32_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<int32_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::i64:
-            replacement = make_range_replacement_integral<int64_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<int64_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::u8:
-            replacement = make_range_replacement_integral<uint8_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<uint8_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::u16:
-            replacement = make_range_replacement_integral<uint16_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<uint16_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::u32:
-            replacement = make_range_replacement_integral<uint32_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<uint32_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::u64:
-            replacement = make_range_replacement_integral<uint64_t>(et, shape, start_arg, step_arg);
+            replacement = make_range_replacement<uint64_t>(et, shape, start_arg, step_arg);
             break;
         case element::Type_t::undefined:
         case element::Type_t::dynamic:

src/ngraph/runtime/reference/range.hpp

+//*****************************************************************************
+// Copyright 2017-2019 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#pragma once
+
+#include <cmath>
+#include <type_traits>
+
+#include "ngraph/axis_vector.hpp"
+#include "ngraph/check.hpp"
+#include "ngraph/coordinate_transform.hpp"
+
+namespace ngraph
+{
+    namespace runtime
+    {
+        namespace reference
+        {
+            // Return type is `void`, only enabled if `T` is a built-in FP
+            // type, or nGraph's `bfloat16` or `float16` type.
+            template <typename T>
+            typename std::enable_if<std::is_floating_point<T>::value ||
+                                    std::is_same<T, bfloat16>::value ||
+                                    std::is_same<T, float16>::value>::type
+                range(const T* start, const T* step, const Shape& out_shape, T* out)
+            {
+                for (size_t i = 0; i < shape_size(out_shape); i++)
+                {
+                    out[i] = *start + (static_cast<T>(i) * (*step));
+                }
+            }
+
+            // Return type is `void`, only enabled if `T` is `is_integral`.
+            template <typename T>
+            typename std::enable_if<std::is_integral<T>::value>::type
+                range(const T* start, const T* step, const Shape& out_shape, T* out)
+            {
+                T val = *start;
+
+                for (size_t i = 0; i < shape_size(out_shape); i++)
+                {
+                    out[i] = val;
+                    val += *step;
+                }
+            }
+        }
+    }
+}

test/constant_folding.cpp

@@ -891,6 +891,94 @@ TEST(constant_folding, constant_transpose)
     ASSERT_TRUE(test::all_close_f(values_permute, values_out, MIN_FLOAT_TOLERANCE_BITS));
 }
 
+void range_test_check(const vector<double>& values_out, const vector<double>& values_expected)
+{
+    ASSERT_TRUE(test::all_close_f(values_out, values_expected, MIN_FLOAT_TOLERANCE_BITS));
+}
+
+void range_test_check(const vector<float>& values_out, const vector<float>& values_expected)
+{
+    ASSERT_TRUE(test::all_close_f(values_out, values_expected, MIN_FLOAT_TOLERANCE_BITS));
+}
+
+template <typename T>
+typename std::enable_if<std::is_integral<T>::value>::type
+    range_test_check(const vector<T>& values_out, const vector<T>& values_expected)
+{
+    ASSERT_EQ(values_out, values_expected);
+}
+
+template <typename T>
+void range_test(T start, T stop, T step, const vector<T>& values_expected)
+{
+    vector<T> values_start{start};
+    vector<T> values_stop{stop};
+    vector<T> values_step{step};
+
+    auto constant_start = make_shared<op::Constant>(element::from<T>(), Shape{}, values_start);
+    auto constant_stop = make_shared<op::Constant>(element::from<T>(), Shape{}, values_stop);
+    auto constant_step = make_shared<op::Constant>(element::from<T>(), Shape{}, values_step);
+    auto range = make_shared<op::Range>(constant_start, constant_stop, constant_step);
+    auto f = make_shared<Function>(range, ParameterVector{});
+
+    pass::Manager pass_manager;
+    pass_manager.register_pass<pass::ConstantFolding>();
+    pass_manager.run_passes(f);
+
+    ASSERT_EQ(count_ops_of_type<op::Range>(f), 0);
+    ASSERT_EQ(count_ops_of_type<op::Constant>(f), 1);
+
+    auto new_const =
+        std::dynamic_pointer_cast<op::Constant>(f->get_results().at(0)->get_argument(0));
+    ASSERT_TRUE(new_const);
+
+    auto values_out = new_const->template get_vector<T>();
+
+    range_test_check(values_out, values_expected);
+}
+
+TEST(constant_folding, constant_range)
+{
+    range_test<int8_t>(5, 12, 2, {5, 7, 9, 11});
+    range_test<int32_t>(5, 12, 2, {5, 7, 9, 11});
+    range_test<int64_t>(5, 12, 2, {5, 7, 9, 11});
+    range_test<uint64_t>(5, 12, 2, {5, 7, 9, 11});
+    range_test<double>(5, 12, 2, {5, 7, 9, 11});
+    range_test<float>(5, 12, 2, {5, 7, 9, 11});
+
+    range_test<int32_t>(5, 12, -2, {});
+    range_test<float>(12, 4, -2, {12, 10, 8, 6});
+}
+
+TEST(constant_folding, constant_select)
+{
+    Shape shape{2, 4};
+    vector<char> values_selection{0, 1, 1, 0, 1, 0, 0, 1};
+    vector<int64_t> values_t{2, 4, 6, 8, 10, 12, 14, 16};
+    vector<int64_t> values_f{1, 3, 5, 7, 9, 11, 13, 15};
+
+    auto constant_selection = make_shared<op::Constant>(element::boolean, shape, values_selection);
+    auto constant_t = make_shared<op::Constant>(element::i64, shape, values_t);
+    auto constant_f = make_shared<op::Constant>(element::i64, shape, values_f);
+    auto select = make_shared<op::Select>(constant_selection, constant_t, constant_f);
+    auto f = make_shared<Function>(select, ParameterVector{});
+
+    pass::Manager pass_manager;
+    pass_manager.register_pass<pass::ConstantFolding>();
+    pass_manager.run_passes(f);
+
+    ASSERT_EQ(count_ops_of_type<op::Select>(f), 0);
+    ASSERT_EQ(count_ops_of_type<op::Constant>(f), 1);
+
+    auto new_const =
+        std::dynamic_pointer_cast<op::Constant>(f->get_results().at(0)->get_argument(0));
+    ASSERT_TRUE(new_const);
+    auto values_out = new_const->get_vector<int64_t>();
+
+    vector<int64_t> values_expected{1, 4, 6, 7, 10, 11, 13, 16};
+    ASSERT_EQ(values_expected, values_out);
+}
+
 TEST(constant_folding, pass_property)
 {
     auto pass = std::make_shared<ngraph::pass::ConstantFolding>();