CPU: Implement op::Reshape

src/ngraph/runtime/cpu/emitter.cpp

@@ -25,6 +25,7 @@
 #include "ngraph/ops/concatenate.hpp"
 #include "ngraph/ops/constant.hpp"
 #include "ngraph/ops/get_tuple_element.hpp"
+#include "ngraph/ops/reshape.hpp"
 #include "ngraph/runtime/cpu/emitter.hpp"
 #include "ngraph/runtime/cpu/external_function.hpp"
 #include "ngraph/runtime/tensor_view_info.hpp"
@@ -931,3 +932,68 @@ void Emitter::EMITTER_DECL(EmitConstant)
 
     TU += "};\n    }\n";
 }
+
+void Emitter::EMITTER_DECL(EmitReshape)
+{
+    auto reshape = static_cast<const op::Reshape*>(n);
+
+    auto arg_type = reshape->get_arguments().at(0)->get_value_type();
+    auto arg_tensor_view_type = dynamic_pointer_cast<const TensorViewType>(arg_type);
+    assert(arg_tensor_view_type);
+    auto arg_shape = arg_tensor_view_type->get_shape();
+    auto arg_rank = arg_shape.size();
+
+    auto result_type = reshape->get_value_type();
+    auto result_tensor_view_type = dynamic_pointer_cast<const TensorViewType>(result_type);
+    assert(result_tensor_view_type);
+    auto result_shape = result_tensor_view_type->get_shape();
+    auto& result_element_type = result_tensor_view_type->get_element_type();
+
+    auto input_order = reshape->get_input_order();
+
+    bool same_layout = std::is_sorted(input_order.begin(), input_order.end());
+
+    size_t result_shape_product = 1;
+    for (auto i : result_shape)
+    {
+        result_shape_product *= i;
+    }
+
+    // If there is no layout change or we are just going from 1^n to 1^m or a zero-size tensor, we can just copy.
+    if (same_layout || result_shape_product < 2)
+    {
+        TU +=
+            "    {\n"
+            "        call_frame->get_parameterized_tensor_view<" +
+            element_type_names[TI(result_element_type)] + ">(" + to_string(outputs.at(0).get_index()) +
+            ")->get_vector() =\n"
+            "        call_frame->get_parameterized_tensor_view<" +
+            element_type_names[TI(result_element_type)] + ">(" +
+            to_string(inputs.at(0).get_index()) +
+            ")->get_vector();\n"
+            "    }\n";
+    }
+    // If there *is* a layout change in the 2D case, we transpose the input.
+    else if (arg_rank == 2)
+    {
+        auto arg0_layout = inputs[0].get_layout<DenseTensorViewLayout>();
+        auto out_layout = outputs[0].get_layout<DenseTensorViewLayout>();
+
+        TU += "    {\n"
+            "        auto arg0 = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
+            ">(" + to_string(inputs[0].get_index()) + ");\n"
+            "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(result_element_type)] +
+            ">(" + to_string(outputs[0].get_index()) + ");\n"
+            "        EigenMatrix<" + element_type_names[TI(result_element_type)] + ">(out, " +
+            EIGEN_MATRIX_FORMAT(out_layout->get_shape(), out_layout->get_strides()) + ") =\n"
+            "        EigenMatrix<" + element_type_names[TI(result_element_type)] + ">(arg0, " +
+            EIGEN_MATRIX_FORMAT(arg0_layout->get_shape(), arg0_layout->get_strides()) + ").transpose();\n"
+            "    }\n";
+    }
+    // Other cases (reordering of axes for tensors with rank>2) are not handled yet.
+    else
+    {
+        throw ngraph_error(
+            "Axis permutation in reshape is not implemented yet for tensors with rank>2");
+    }
+}

src/ngraph/runtime/cpu/emitter.hpp

@@ -76,6 +76,7 @@ namespace ngraph
                 void EMITTER_DECL(EmitBroadcast);
                 void EMITTER_DECL(EmitConvert);
                 void EMITTER_DECL(EmitConstant);
+                void EMITTER_DECL(EmitReshape);
             };
         }
     }

src/ngraph/runtime/cpu/external_function.cpp

@@ -47,6 +47,7 @@
 #include "ngraph/ops/negative.hpp"
 #include "ngraph/ops/not_equal.hpp"
 #include "ngraph/ops/reduce.hpp"
+#include "ngraph/ops/reshape.hpp"
 #include "ngraph/ops/select.hpp"
 #include "ngraph/ops/subtract.hpp"
 #include "ngraph/ops/tuple.hpp"
@@ -106,6 +107,7 @@ static const OpMap dispatcher{
     {TI(ngraph::op::Broadcast), &Emitter::EmitBroadcast},
     {TI(ngraph::op::Convert), &Emitter::EmitConvert},
     {TI(ngraph::op::Constant), &Emitter::EmitConstant},
+    {TI(ngraph::op::Reshape), &Emitter::EmitReshape},
 };
 
 #undef TI

test/cpu.cpp

@@ -1655,8 +1655,9 @@ TEST(execute, reduce_matrix_to_scalar_zero_by_zero)
     ASSERT_EQ((vector<float>{}), a->get_vector());
     ASSERT_EQ((vector<float>{99}), b->get_vector());
 }
+*/
 
-TEST(execute, reshape_t2v_012)
+TEST(cpu, reshape_t2v_012)
 {
     auto shape_a = Shape{2, 2, 3};
     auto A = make_shared<op::Parameter>(
@@ -1666,7 +1667,7 @@ TEST(execute, reshape_t2v_012)
     auto r = make_shared<op::Reshape>(A, AxisVector{0, 1, 2}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1680,7 +1681,7 @@ TEST(execute, reshape_t2v_012)
     ASSERT_EQ((vector<float>{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}), result->get_vector());
 }
 
-TEST(execute, reshape_t2s_012)
+TEST(cpu, reshape_t2s_012)
 {
     auto shape_a = Shape{1, 1, 1};
     auto A = make_shared<op::Parameter>(
@@ -1690,7 +1691,7 @@ TEST(execute, reshape_t2s_012)
     auto r = make_shared<op::Reshape>(A, AxisVector{0, 1, 2}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1704,7 +1705,7 @@ TEST(execute, reshape_t2s_012)
     ASSERT_EQ((vector<float>{6}), result->get_vector());
 }
 
-TEST(execute, reshape_t2s_120)
+TEST(cpu, reshape_t2s_120)
 {
     auto shape_a = Shape{1, 1, 1};
     auto A = make_shared<op::Parameter>(
@@ -1714,7 +1715,7 @@ TEST(execute, reshape_t2s_120)
     auto r = make_shared<op::Reshape>(A, AxisVector{1, 2, 0}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1728,7 +1729,7 @@ TEST(execute, reshape_t2s_120)
     ASSERT_EQ((vector<float>{6}), result->get_vector());
 }
 
-TEST(execute, reshape_s2t)
+TEST(cpu, reshape_s2t)
 {
     auto shape_a = Shape{};
     auto A = make_shared<op::Parameter>(
@@ -1738,7 +1739,7 @@ TEST(execute, reshape_s2t)
     auto r = make_shared<op::Reshape>(A, AxisVector{}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1752,7 +1753,7 @@ TEST(execute, reshape_s2t)
     ASSERT_EQ((vector<float>{42}), result->get_vector());
 }
 
-TEST(execute, reshape_v2m_col)
+TEST(cpu, reshape_v2m_col)
 {
     auto shape_a = Shape{3};
     auto A = make_shared<op::Parameter>(
@@ -1762,7 +1763,7 @@ TEST(execute, reshape_v2m_col)
     auto r = make_shared<op::Reshape>(A, AxisVector{0}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1776,7 +1777,7 @@ TEST(execute, reshape_v2m_col)
     ASSERT_EQ((vector<float>{1, 2, 3}), result->get_vector());
 }
 
-TEST(execute, reshape_v2m_row)
+TEST(cpu, reshape_v2m_row)
 {
     auto shape_a = Shape{3};
     auto A = make_shared<op::Parameter>(
@@ -1786,7 +1787,7 @@ TEST(execute, reshape_v2m_row)
     auto r = make_shared<op::Reshape>(A, AxisVector{0}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1800,7 +1801,7 @@ TEST(execute, reshape_v2m_row)
     ASSERT_EQ((vector<float>{1, 2, 3}), result->get_vector());
 }
 
-TEST(execute, reshape_v2t_middle)
+TEST(cpu, reshape_v2t_middle)
 {
     auto shape_a = Shape{3};
     auto A = make_shared<op::Parameter>(
@@ -1810,7 +1811,7 @@ TEST(execute, reshape_v2t_middle)
     auto r = make_shared<op::Reshape>(A, AxisVector{0}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1824,7 +1825,7 @@ TEST(execute, reshape_v2t_middle)
     ASSERT_EQ((vector<float>{1, 2, 3}), result->get_vector());
 }
 
-TEST(execute, reshape_m2m_same)
+TEST(cpu, reshape_m2m_same)
 {
     auto shape_a = Shape{3, 3};
     auto A = make_shared<op::Parameter>(
@@ -1834,7 +1835,7 @@ TEST(execute, reshape_m2m_same)
     auto r = make_shared<op::Reshape>(A, AxisVector{0, 1}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1848,7 +1849,7 @@ TEST(execute, reshape_m2m_same)
     ASSERT_EQ((vector<float>{1, 2, 3, 4, 5, 6, 7, 8, 9}), result->get_vector());
 }
 
-TEST(execute, reshape_m2m_transpose)
+TEST(cpu, reshape_m2m_transpose)
 {
     auto shape_a = Shape{3, 3};
     auto A = make_shared<op::Parameter>(
@@ -1858,7 +1859,7 @@ TEST(execute, reshape_m2m_transpose)
     auto r = make_shared<op::Reshape>(A, AxisVector{1, 0}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1872,7 +1873,7 @@ TEST(execute, reshape_m2m_transpose)
     ASSERT_EQ((vector<float>{1, 4, 7, 2, 5, 8, 3, 6, 9}), result->get_vector());
 }
 
-TEST(execute, reshape_m2m_dim_change_transpose)
+TEST(cpu, reshape_m2m_dim_change_transpose)
 {
     auto shape_a = Shape{3, 2};
     auto A = make_shared<op::Parameter>(
@@ -1882,7 +1883,7 @@ TEST(execute, reshape_m2m_dim_change_transpose)
     auto r = make_shared<op::Reshape>(A, AxisVector{1, 0}, shape_r);
     auto f = make_shared<Function>(r, rt, op::Parameters{A});
 
-    auto manager = runtime::Manager::get("NGVM");
+    auto manager = runtime::Manager::get("CPU");
     auto external = manager->compile(f);
     auto backend = manager->allocate_backend();
     auto cf = backend->make_call_frame(external);
@@ -1896,6 +1897,7 @@ TEST(execute, reshape_m2m_dim_change_transpose)
     ASSERT_EQ((vector<float>{1, 3, 5, 2, 4, 6}), result->get_vector());
 }
 
+/*
 TEST(execute, sin)
 {
     auto shape = Shape{6};