[Snippets] Added Softmax support

src/common/snippets/include/snippets/op/buffer.hpp

@@ -0,0 +1,42 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <ngraph/op/op.hpp>
+
+namespace ngraph {
+namespace snippets {
+namespace op {
+
+/**
+ * @interface Buffer
+ * @brief TODO
+ * @ingroup snippets
+ */
+class Buffer : public ngraph::op::Op {
+public:
+    OPENVINO_OP("Buffer", "SnippetsOpset");
+
+    Buffer(const Output<Node>& x, const size_t offset = 0);
+    Buffer() = default;
+
+    size_t get_offset() const;
+    void set_offset(const size_t offset);
+
+    // If Buffer has offset this method set this offset to near Load and Store ops
+    // to correctly read and write data
+    void propogateOffset();
+
+    bool visit_attributes(AttributeVisitor& visitor) override { return true; };
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+    void validate_and_infer_types() override;
+
+private:
+    size_t offset;
+};
+
+} // namespace op
+} // namespace snippets
+} // namespace ngraph

src/common/snippets/include/snippets/op/fill.hpp

@@ -0,0 +1,44 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <ngraph/op/op.hpp>
+
+namespace ngraph {
+namespace snippets {
+namespace op {
+
+/**
+ * @interface Fill
+ * @brief TODO
+ * @ingroup snippets
+ */
+class Fill : public ngraph::op::Op {
+public:
+    OPENVINO_OP("Fill", "SnippetsOpset");
+
+    Fill(const Output<Node>& x, const int64_t offset, const std::string fill_value = "zero");
+    Fill() = default;
+
+    int64_t get_offset() const { return m_offset; }
+    std::string get_fill_value() const { return m_fill_value; }
+
+    void set_offset(const size_t offset) { m_offset = offset; }
+    void set_fill_value(const std::string fill_value) { m_fill_value = fill_value; }
+
+    bool visit_attributes(AttributeVisitor& visitor) override;
+
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+
+    void validate_and_infer_types() override;
+
+protected:
+    int64_t m_offset = 0lu;
+    std::string m_fill_value = "zero";
+};
+
+} // namespace op
+} // namespace snippets
+} // namespace ngraph

src/common/snippets/include/snippets/op/horizon_max.hpp

@@ -0,0 +1,33 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/op/op.hpp"
+
+namespace ngraph {
+namespace snippets {
+namespace op {
+
+/**
+ * @interface HorizonMax
+ * @brief TODO
+ * @ingroup snippets
+ */
+class HorizonMax : public ngraph::op::Op {
+public:
+    OPENVINO_OP("HorizonMax", "SnippetsOpset");
+
+    HorizonMax(const Output<Node>& x);
+    HorizonMax() = default;
+
+    bool visit_attributes(AttributeVisitor& visitor) override { return true;}
+
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+    void validate_and_infer_types() override;
+};
+
+} // namespace op
+} // namespace snippets
+} // namespace ngraph

src/common/snippets/include/snippets/op/horizon_sum.hpp

@@ -0,0 +1,32 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "ngraph/op/op.hpp"
+
+namespace ngraph {
+namespace snippets {
+namespace op {
+
+/**
+ * @interface HorizonSum
+ * @brief TODO
+ * @ingroup snippets
+ */
+class HorizonSum : public ngraph::op::Op {
+public:
+    OPENVINO_OP("HorizonSum", "SnippetsOpset");
+
+    HorizonSum(const Output<Node>& x);
+    HorizonSum() = default;
+
+    bool visit_attributes(AttributeVisitor& visitor) override { return true;}
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+    void validate_and_infer_types() override;
+};
+
+} // namespace op
+} // namespace snippets
+} // namespace ngraph

src/common/snippets/include/snippets/op/load.hpp

@@ -15,18 +15,21 @@ namespace op {
  * @brief Generated by Canonicalization step where explicit instructions should be emitted for data loading
  *        where number of elements to load is determined by "count"
  *        Default value is "1" - to load one element
+ *        TODO: ADD DESCRIPTION OF PARAMS
  * @ingroup snippets
  */
 class Load : public ngraph::op::Op {
 public:
     OPENVINO_OP("Load", "SnippetsOpset");
 
-    Load(const Output<Node>& x, const size_t count = 1lu);
+    Load(const Output<Node>& x, const size_t count = 1lu, const size_t offset = 0lu);
     Load() = default;
 
     size_t get_count() const { return m_count; }
+    size_t get_offset() const { return m_offset; }
 
     void set_count(const size_t count) { m_count = count; }
+    void set_offset(const size_t offset) { m_offset = offset; }
 
     bool visit_attributes(AttributeVisitor& visitor) override;
 
@@ -40,6 +43,7 @@ class Load : public ngraph::op::Op {
 
 protected:
     size_t m_count = 0lu;
+    size_t m_offset = 0lu;
 };
 
 } // namespace op

src/common/snippets/include/snippets/op/store.hpp

@@ -21,12 +21,14 @@ class Store : public ngraph::op::Op {
 public:
     OPENVINO_OP("Store", "SnippetsOpset");
 
-    Store(const Output<Node>& x, const size_t count = 1lu);
+    Store(const Output<Node>& x, const size_t count = 1lu, const size_t offset = 0lu);
     Store() = default;
 
     size_t get_count() const { return m_count; }
+    size_t get_offset() const { return m_offset; }
 
     void set_count(const size_t count) { m_count = count; }
+    void set_offset(const size_t offset) { m_offset = offset; }
 
     bool visit_attributes(AttributeVisitor& visitor) override;
 
@@ -40,6 +42,7 @@ class Store : public ngraph::op::Op {
 
 protected:
     size_t m_count = 0lu;
+    size_t m_offset = 0lu;
 };
 
 } // namespace op

src/common/snippets/include/snippets/op/subgraph.hpp

@@ -89,17 +89,10 @@ class Subgraph : public ngraph::op::Op {
         return m_generator;
     }
 
-    size_t get_non_scalar_constants_count() const {
-        return m_non_scalar_constants_count;
-    }
-
-    bool is_quantized() const {
-        return config.m_is_quantized;
-    }
-
-    bool has_type_relaxed_ops() const {
-        return config.m_has_type_relaxed_ops;
-    }
+    size_t get_buffer_scratchpad_size() const;
+    size_t get_additional_data_count() const { return m_additional_data_count; }
+    bool is_quantized() const { return config.m_is_quantized; }
+    bool has_type_relaxed_ops() const { return config.m_has_type_relaxed_ops; }
 
     snippets::Schedule generate(const BlockedShapeVector& output_shapes, const BlockedShapeVector& input_shapes, ngraph::pass::Manager& opt,
                                 const void* compile_params = nullptr);
@@ -114,7 +107,7 @@ class Subgraph : public ngraph::op::Op {
     // plugin sets generator for a snippet to some specific generator.
     // it's going to be replaced with Jitters table later
     void set_generator(std::shared_ptr<ngraph::snippets::Generator> generator);
-    void set_non_scalar_constants_count(const size_t count);
+    void set_additional_data_count(const size_t count);
 
     void print() const;
     void print_statistics(bool verbose);
@@ -128,11 +121,12 @@ class Subgraph : public ngraph::op::Op {
 private:
     void align_element_types(const BlockedShapeVector& outputShapes, const BlockedShapeVector& inputShapes);
     void convert_to_snippet_dialect();
+    void init_config();
     // Count of potentional non-scalar Consants that will be created after some tranformations
     // At the moment it's relevant only for FakeQuantize decomposition
     // NOTE: To avoid overheads in each calcution of this count (for example, in validate_and_type_infer()),
     //       we should MANUALLY calculate it where it needed.
-    size_t m_non_scalar_constants_count = 0;
+    size_t m_additional_data_count = 0;
     Shape exec_domain = {};
     std::shared_ptr<ov::Model> m_body = nullptr;
     std::shared_ptr<ngraph::snippets::Generator> m_generator = nullptr;
@@ -150,6 +144,8 @@ class Subgraph : public ngraph::op::Op {
         // True if Subgraph contains TypeRelaxed nodes -> for several streams in tp mode we should copy body using mutexes
         // because TypeRelaxed::copy_with_new_inputs() isn't save-thread method
         bool m_has_type_relaxed_ops = false;
+        // True if we should check runtime info for nodes to call specific needed transformations
+        bool m_check_rt_info = false;
     } config;
 
     ov::PartialShape master_shape;

src/common/snippets/include/snippets/op/vector_buffer.hpp

@@ -0,0 +1,31 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <ngraph/op/op.hpp>
+
+namespace ngraph {
+namespace snippets {
+namespace op {
+
+/**
+ * @interface Buffer
+ * @brief TODO
+ * @ingroup snippets
+ */
+class VectorBuffer : public ngraph::op::Op {
+public:
+    OPENVINO_OP("VectorBuffer", "SnippetsOpset");
+
+    VectorBuffer();
+
+    bool visit_attributes(AttributeVisitor& visitor) override { return true;}
+    std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& new_args) const override;
+    void validate_and_infer_types() override;
+};
+
+} // namespace op
+} // namespace snippets
+} // namespace ngraph

src/common/snippets/include/snippets/pass/set_buffer_offset.hpp

@@ -0,0 +1,31 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <ngraph/pass/graph_rewrite.hpp>
+#include <ngraph/pattern/matcher.hpp>
+
+namespace ngraph {
+namespace snippets {
+namespace pass {
+
+/**
+ * @interface SetBufferOffset
+ * @brief TODO
+ *        NOTE: Should be called after Load/Store insertion and before LoadMoveBroadcastToBroadcastLoad because
+ *              we cannot fuse Load with non-zero offset and MoveBroadcast
+ * @ingroup snippets
+ */
+class SetBufferOffset: public ngraph::pass::MatcherPass {
+public:
+    SetBufferOffset();
+
+private:
+    size_t current_offset = 0lu;
+};
+
+} // namespace pass
+} // namespace snippets
+} // namespace ngraph

src/common/snippets/include/snippets/pass/softmax_decomposition.hpp

@@ -0,0 +1,26 @@
+// Copyright (C) 2018-2022 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include <ngraph/pass/graph_rewrite.hpp>
+#include <ngraph/pattern/matcher.hpp>
+
+namespace ngraph {
+namespace snippets {
+namespace pass {
+
+/**
+ * @interface SoftmaxDecomposition
+ * @brief TODO
+ * @ingroup snippets
+ */
+class SoftmaxDecomposition: public ngraph::pass::MatcherPass {
+public:
+    SoftmaxDecomposition(const size_t vector_size);
+};
+
+}  // namespace pass
+}  // namespace snippets
+}  // namespace ngraph

src/common/snippets/include/snippets/snippets_isa.hpp

@@ -9,15 +9,20 @@
 
 #include "op/broadcastload.hpp"
 #include "op/broadcastmove.hpp"
+#include "op/buffer.hpp"
 #include "op/convert_saturation.hpp"
 #include "op/convert_truncation.hpp"
+#include "op/horizon_max.hpp"
+#include "op/horizon_sum.hpp"
+#include "op/fill.hpp"
 #include "op/kernel.hpp"
 #include "op/load.hpp"
 #include "op/nop.hpp"
 #include "op/scalar.hpp"
 #include "op/powerstatic.hpp"
 #include "op/store.hpp"
 #include "op/loop.hpp"
+#include "op/vector_buffer.hpp"
 
 namespace ngraph {
 namespace snippets {