run Quantize / Dequantize reference kernels on CPU backend (#1780)

* dex core quantize/de

* more types

* codegen

* remove comments

* remove more dead code

src/ngraph/runtime/cpu/CMakeLists.txt

@@ -70,6 +70,7 @@ set(SRC
     builder/softmax.cpp
     builder/sum.cpp
     builder/topk.cpp
+    builder/quantization.cpp
     kernel/eigen_thread_pool.cpp
     kernel/pad.cpp
     kernel/reduce_max.cpp

src/ngraph/runtime/cpu/cpu_emitter.cpp

@@ -42,6 +42,7 @@
 #include "ngraph/op/convolution.hpp"
 #include "ngraph/op/cos.hpp"
 #include "ngraph/op/cosh.hpp"
+#include "ngraph/op/dequantize.hpp"
 #include "ngraph/op/divide.hpp"
 #include "ngraph/op/dot.hpp"
 #include "ngraph/op/equal.hpp"
@@ -71,6 +72,7 @@
 #include "ngraph/op/parameter.hpp"
 #include "ngraph/op/power.hpp"
 #include "ngraph/op/product.hpp"
+#include "ngraph/op/quantize.hpp"
 #include "ngraph/op/reduce.hpp"
 #include "ngraph/op/reduce_window.hpp"
 #include "ngraph/op/relu.hpp"
@@ -4690,6 +4692,34 @@ namespace ngraph
                 writer.block_end();
             }
 
+            template <>
+            void CPU_Emitter::EMITTER_DECL(ngraph::op::Dequantize)
+            {
+                auto dequantize = static_cast<const ngraph::op::Dequantize*>(node);
+                writer << "reference::dequantize(";
+                writer << "            " << args[0].get_name() << ",\n";
+                writer << "            " << args[1].get_name() << ",\n";
+                writer << "            " << args[2].get_name() << ",\n";
+                writer << "            " << out[0].get_name() << ",\n";
+                writer << "            {" << join(args[0].get_shape()) << "},\n";
+                writer << "            {" << join(args[1].get_shape()) << "},\n";
+                writer << "            {" << join(dequantize->get_axes()) << "});\n";
+            }
+
+            template <>
+            void CPU_Emitter::EMITTER_DECL(ngraph::op::Quantize)
+            {
+                auto quantize = static_cast<const ngraph::op::Dequantize*>(node);
+                writer << "reference::quantize(";
+                writer << "            " << args[0].get_name() << ",\n";
+                writer << "            " << args[1].get_name() << ",\n";
+                writer << "            " << args[2].get_name() << ",\n";
+                writer << "            " << out[0].get_name() << ",\n";
+                writer << "            {" << join(args[0].get_shape()) << "},\n";
+                writer << "            {" << join(args[1].get_shape()) << "},\n";
+                writer << "            {" << join(quantize->get_axes()) << "});\n";
+            }
+
 #undef TI
         }
     }

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -65,6 +65,7 @@
 #include "ngraph/op/convolution.hpp"
 #include "ngraph/op/cos.hpp"
 #include "ngraph/op/cosh.hpp"
+#include "ngraph/op/dequantize.hpp"
 #include "ngraph/op/divide.hpp"
 #include "ngraph/op/dot.hpp"
 #include "ngraph/op/equal.hpp"
@@ -94,6 +95,7 @@
 #include "ngraph/op/parameter.hpp"
 #include "ngraph/op/power.hpp"
 #include "ngraph/op/product.hpp"
+#include "ngraph/op/quantize.hpp"
 #include "ngraph/op/reduce.hpp"
 #include "ngraph/op/reduce_window.hpp"
 #include "ngraph/op/relu.hpp"
@@ -362,6 +364,8 @@ static const runtime::cpu::OpMap dispatcher{
      &runtime::cpu::CPU_Emitter::emit<runtime::cpu::op::LoopKernel>},
     {TI(ngraph::op::LRN), &runtime::cpu::CPU_Emitter::emit<ngraph::op::LRN>},
     {TI(ngraph::op::ConvolutionAdd), &runtime::cpu::CPU_Emitter::emit<op::ConvolutionAdd>},
+    {TI(ngraph::op::Quantize), &runtime::cpu::CPU_Emitter::emit<op::Quantize>},
+    {TI(ngraph::op::Dequantize), &runtime::cpu::CPU_Emitter::emit<op::Dequantize>},
 
 };
 
@@ -436,6 +440,7 @@ void runtime::cpu::CPU_ExternalFunction::compile()
 #include "ngraph/runtime/reference/broadcast.hpp"
 #include "ngraph/runtime/reference/concat.hpp"
 #include "ngraph/runtime/reference/convolution.hpp"
+#include "ngraph/runtime/reference/dequantize.hpp"
 #include "ngraph/runtime/reference/dot.hpp"
 #include "ngraph/runtime/reference/lrn.hpp"
 #include "ngraph/runtime/reference/max.hpp"
@@ -446,6 +451,7 @@ void runtime::cpu::CPU_ExternalFunction::compile()
 #include "ngraph/runtime/reference/or.hpp"
 #include "ngraph/runtime/reference/pad.hpp"
 #include "ngraph/runtime/reference/product.hpp"
+#include "ngraph/runtime/reference/quantize.hpp"
 #include "ngraph/runtime/reference/reduce.hpp"
 #include "ngraph/runtime/reference/reduce_window.hpp"
 #include "ngraph/runtime/reference/relu.hpp"

src/ngraph/runtime/cpu/unit_test.manifest

@@ -12,10 +12,3 @@ one_hot_vector_1_fp
 one_hot_vector_1_fp_nonint
 backwards_batch_norm_three_outputs
 backwards_maxpool_n2_c1_hw5_3x3_str2_max_pad1x2_2x3
-quantize
-quantize_axes
-quantize_int8
-quantize_clamp
-dequantize
-dequantize_axes
-dequantize_int8