[CPU] [DEBUG CAPS] Extension for snippets and other ngraph transformations (#14223)

Author: EgorDuplensky

src/plugins/intel_cpu/src/config.cpp

@@ -14,10 +14,11 @@
 #include "ie_parallel.hpp"
 #include "ie_system_conf.h"
 
-#include <cpp_interfaces/interface/ie_internal_plugin_config.hpp>
+#include "cpp_interfaces/interface/ie_internal_plugin_config.hpp"
 #include "openvino/core/type/element_type_traits.hpp"
 #include "openvino/runtime/properties.hpp"
-#include <cpu/x64/cpu_isa_traits.hpp>
+#include "utils/debug_capabilities.h"
+#include "cpu/x64/cpu_isa_traits.hpp"
 
 namespace ov {
 namespace intel_cpu {
@@ -48,10 +49,24 @@ Config::Config() {
     if (!dnnl::impl::cpu::x64::mayiuse(dnnl::impl::cpu::x64::avx512_core_bf16))
         enforceBF16 = false;
 
-    CPU_DEBUG_CAP_ENABLE(readDebugCapsProperties());
+    CPU_DEBUG_CAP_ENABLE(applyDebugCapsProperties());
+
     updateProperties();
 }
 
+#ifdef CPU_DEBUG_CAPS
+/**
+ * Debug capabilities configuration has more priority than common one
+ * Some of the debug capabilities also require to enable some of common
+ * configuration properties
+ */
+void Config::applyDebugCapsProperties() {
+    // always enable perf counters for verbose mode and performance summary
+    if (!debugCaps.verbose.empty() || !debugCaps.summaryPerf.empty())
+        collectPerfCounters = true;
+}
+#endif
+
 void Config::readProperties(const std::map<std::string, std::string> &prop) {
     const auto streamExecutorConfigKeys = streamExecutorConfig.SupportedKeys();
     const auto hintsConfigKeys = perfHintsConfig.SupportedKeys();
@@ -184,7 +199,7 @@ void Config::readProperties(const std::map<std::string, std::string> &prop) {
     if (exclusiveAsyncRequests)  // Exclusive request feature disables the streams
         streamExecutorConfig._streams = 1;
 
-    CPU_DEBUG_CAP_ENABLE(readDebugCapsProperties());
+    CPU_DEBUG_CAP_ENABLE(applyDebugCapsProperties());
     updateProperties();
 }
 
@@ -239,58 +254,6 @@ void Config::updateProperties() {
     _config.insert({PluginConfigParams::KEY_CACHE_DIR, cache_dir});
 }
 
-#ifdef CPU_DEBUG_CAPS
-void Config::readDebugCapsProperties() {
-    auto readEnv = [](const char* envVar) {
-        return std::getenv(envVar);
-    };
-
-    auto parseDumpFormat = [](const std::string& format) {
-        if (format == "BIN")
-            return FORMAT::BIN;
-        else if (format == "TEXT")
-            return FORMAT::TEXT;
-        else
-            IE_THROW() << "readDebugCapsProperties: Unknown dump format";
-    };
-
-    const char* envVarValue = nullptr;
-
-    if (envVarValue = readEnv("OV_CPU_EXEC_GRAPH_PATH"))
-        execGraphPath = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_VERBOSE"))
-        verbose = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_BLOB_DUMP_DIR"))
-        blobDumpDir = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_BLOB_DUMP_FORMAT"))
-        blobDumpFormat = parseDumpFormat(envVarValue);
-
-    if (envVarValue = readEnv("OV_CPU_BLOB_DUMP_NODE_EXEC_ID"))
-        blobDumpFilters[BY_EXEC_ID] = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_BLOB_DUMP_NODE_PORTS"))
-        blobDumpFilters[BY_PORTS] = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_BLOB_DUMP_NODE_TYPE"))
-        blobDumpFilters[BY_TYPE] = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_BLOB_DUMP_NODE_NAME"))
-        blobDumpFilters[BY_NAME] = envVarValue;
-
-    if (envVarValue = readEnv("OV_CPU_SUMMARY_PERF")) {
-        collectPerfCounters = true;
-        summaryPerf = envVarValue;
-    }
-
-    // always enable perf counters for verbose mode
-    if (!verbose.empty())
-        collectPerfCounters = true;
-}
-#endif // CPU_DEBUG_CAPS
-
 }   // namespace intel_cpu
 }   // namespace ov
 

src/plugins/intel_cpu/src/config.h

@@ -6,8 +6,11 @@
 
 #include <threading/ie_istreams_executor.hpp>
 #include <ie_performance_hints.hpp>
-#include "utils/debug_capabilities.h"
+#include <ie/ie_common.h>
+#include <openvino/util/common_util.hpp>
+#include "utils/debug_caps_config.h"
 
+#include <bitset>
 #include <string>
 #include <map>
 #include <mutex>
@@ -57,31 +60,12 @@ struct Config {
 
     std::map<std::string, std::string> _config;
 
-#ifdef CPU_DEBUG_CAPS
-    enum FILTER {
-        BY_PORTS,
-        BY_EXEC_ID,
-        BY_TYPE,
-        BY_NAME,
-    };
-
-    enum class FORMAT {
-        BIN,
-        TEXT,
-    };
-
-    std::string execGraphPath;
-    std::string verbose;
-    std::string blobDumpDir = "cpu_dump";
-    FORMAT blobDumpFormat = FORMAT::TEXT;
-    // std::hash<int> is necessary for Ubuntu-16.04 (gcc-5.4 and defect in C++11 standart)
-    std::unordered_map<FILTER, std::string, std::hash<int>> blobDumpFilters;
-    std::string summaryPerf = "";
+    bool isNewApi = true;
 
-    void readDebugCapsProperties();
+#ifdef CPU_DEBUG_CAPS
+    DebugCapsConfig debugCaps;
+    void applyDebugCapsProperties();
 #endif
-
-    bool isNewApi = true;
 };
 
 }   // namespace intel_cpu

src/plugins/intel_cpu/src/docs/README.md

@@ -6,6 +6,7 @@ Use the following cmake option to enable debug capabilities:
 * [Verbose mode](verbose.md)
 * [Blob dumping](blob_dumping.md)
 * [Graph serialization](graph_serialization.md)
+* [Graph transformation disabling](feature_disabling.md#graph-transformations)
 
 ## Debug log
 

src/plugins/intel_cpu/src/docs/blob_dumping.md

@@ -29,8 +29,8 @@ Default is *cpu_dump*
     OV_CPU_BLOB_DUMP_FORMAT=<format> binary ...
 ```
 Options are:
-* BIN (default)
-* TEXT
+* BIN
+* TEXT (default)
 
 ## Filter input / output blobs
 To dump only input / output blobs:

src/plugins/intel_cpu/src/docs/debug_caps_filters.md

@@ -0,0 +1,52 @@
+# Filters
+
+Filters described below have the following common format:
+```sh
+    filter_name=<comma_separated_tokens>
+```
+Tokens are processed from left to right and each one includes or excludes corresponding value.\
+For exclusion token is just prepended by minus: *-token*\
+All tokens are case insensitive and no tokens is treated as *all*\
+So filters below are equal:
+* filter_name
+* filter_name=all
+* filter_name=-all,ALL
+
+## IR format filter
+
+IR format filter is used to specify output IR formats, e.g. for [serialization](graph_serialization.md#graph-transformations).
+```sh
+    formats=<comma_separated_tokens>
+```
+
+The following tokens are supported:
+* all\
+equals to <xml,dot,svg>
+* xml (default)\
+IR in .xml file. Can be opened using, for example, *netron* app. (For now the option is Linux only)
+* xmlbin\
+IR in .xml and .bin files. Can be opened using, for example, *netron* app.
+* dot\
+IR in .dot file (.svg.dot file if svg is also specified). Can be inspected using, for example, *graphviz* tools.
+* svg\
+IR in .svg file. Requires *dot* tool to be installed on the host, not supported on Windows.\
+Generation is based on dot representation, so IR is additionally dumped to .svg.dot file.
+
+## Transformation filter
+
+Transformation filter is used to specify main graph transformation stages for different purposes,
+e.g. for [disabling](feature_disabling.md#graph-transformation) or [serialization](graph_serialization.md#graph-transformations).
+```sh
+    transformations=<comma_separated_tokens>
+```
+
+The following tokens are supported:
+* all (default)\
+equals to <preLpt,lpt,postLpt,snippets,specific>
+* common \
+equals to <preLpt,postLpt>
+* preLpt
+* lpt
+* postLpt
+* snippets
+* specific

src/plugins/intel_cpu/src/docs/feature_disabling.md

@@ -0,0 +1,26 @@
+# Feature disabling
+
+Common way to disable something in CPU plugin is implied by means of environment variable **OV_CPU_DISABLE**:
+```sh
+    OV_CPU_DISABLE=<space_separated_options> binary ...
+```
+Option names are case insensitive and processed from left to right,\
+so last one overwrites previous ones if duplicated.
+
+Examples:
+```sh
+    OV_CPU_DISABLE="transformations" binary ...
+    OV_CPU_DISABLE="transformations=lpt" binary ...
+    OV_CPU_DISABLE="transformations=all,-common" binary ...
+```
+
+By means of corresponding options **OV_CPU_DISABLE** controls disabling of the following features:
+
+## Graph transformations
+
+Graph transformation disabling is controlled by the following option inside **OV_CPU_DISABLE**:
+```sh
+transformations=<comma_separated_tokens>
+```
+Filter with main transformation stages to disable specified ones.\
+See [transformation filter](debug_caps_filters.md#transformation-filter) for more details.

src/plugins/intel_cpu/src/docs/graph_serialization.md

@@ -1,17 +1,43 @@
 # Graph serialization
 
-The functionality allows to serialize execution graph using environment variable:
+Graph serialization is disabled by default and controlled by environment variables.
+
+## Execution graph
+
+Execution graph could be serialized using environment variable **OV_CPU_EXEC_GRAPH_PATH**:
 ```sh
-    OV_CPU_EXEC_GRAPH_PATH=<path> binary ...
+    OV_CPU_EXEC_GRAPH_PATH=<option> binary ...
 ```
-
 Possible serialization options:
-* cout
+* cout\
+Serialize to console output.
+* \<path\>.xml\
+Serialize graph into .xml and .bin files. Can be opened using, for example, *netron* app.
+* **TBD**: \<path\>.dot\
+Serialize graph into .dot file. Can be inspected using, for example, *graphviz* tools.
 
-    Serialize to console output
-* \<path\>.xml
+## Graph transformations
+
+Additionally, IR could be serialized at specified stages using environment variable **OV_CPU_DUMP_IR**:
+```sh
+    OV_CPU_DUMP_IR=<space_separated_options> binary ...
+```
+
+Examples:
+```sh
+    OV_CPU_DUMP_IR="transformations" binary ...
+    OV_CPU_DUMP_IR="transformations=snippets dir=path/dumpDir" binary ...
+    OV_CPU_DUMP_IR="transformations=all,-common DIR=path/dumpdir formats=svg,xml" binary ...
+```
 
-    Serialize graph into .xml and .bin files. Can be opened using, for example, *netron* app
-* \<path\>.dot
+Option names are case insensitive, the following options are supported:
+* dir=\<path\>\
+Path to dumped IR files. If omitted, it defaults to *intel_cpu_dump*
+* formats=<comma_separated_tokens>\
+Filter with IR formats to dump. If omitted, it defaults to *xml*\
+See [IR format filter](debug_caps_filters.md#ir-format-filter) for more details.
+* transformations=<comma_separated_tokens>\
+Filter with main transformation stages to serialize graph before and after specified ones.\
+See [transformation filter](debug_caps_filters.md#transformation-filter) for more details.
 
-    TBD. Serialize graph into .dot file. Can be inspected using, for example, *graphviz* tools.
+Options are processed from left to right, so last one overwrites previous ones if duplicated.

src/plugins/intel_cpu/src/graph.cpp

@@ -1073,7 +1073,7 @@ void Graph::InferStatic(InferRequestBase* request) {
     dnnl::stream stream(eng);
 
     for (const auto& node : executableGraphNodes) {
-        VERBOSE(node, config.verbose);
+        VERBOSE(node, config.debugCaps.verbose);
         PERF(node, config.collectPerfCounters);
 
         if (request)
@@ -1160,7 +1160,7 @@ void Graph::InferDynamic(InferRequestBase* request) {
         updateNodes(stopIndx);
         for (; inferCounter < stopIndx; ++inferCounter) {
             auto& node = executableGraphNodes[inferCounter];
-            VERBOSE(node, config.verbose);
+            VERBOSE(node, config.debugCaps.verbose);
             PERF(node, config.collectPerfCounters);
 
             if (request)
@@ -1171,7 +1171,7 @@ void Graph::InferDynamic(InferRequestBase* request) {
 }
 
 inline void Graph::ExecuteNode(const NodePtr& node, const dnnl::stream& stream) const {
-    DUMP(node, config, infer_count);
+    DUMP(node, config.debugCaps, infer_count);
     OV_ITT_SCOPED_TASK(itt::domains::intel_cpu, node->profiling.execute);
 
     if (node->isDynamicNode()) {

src/plugins/intel_cpu/src/graph_dumper.cpp

@@ -210,7 +210,7 @@ std::shared_ptr<ngraph::Function> dump_graph_as_ie_ngraph_net(const Graph &graph
 
 #ifdef CPU_DEBUG_CAPS
 void serialize(const Graph &graph) {
-    const std::string& path = graph.getConfig().execGraphPath;
+    const std::string& path = graph.getConfig().debugCaps.execGraphPath;
 
     if (path.empty())
         return;
@@ -257,7 +257,7 @@ void serializeToCout(const Graph &graph) {
 }
 
 void summary_perf(const Graph &graph) {
-    const std::string& summaryPerf = graph.getConfig().summaryPerf;
+    const std::string& summaryPerf = graph.getConfig().debugCaps.summaryPerf;
 
     if (summaryPerf.empty())
         return;

src/plugins/intel_cpu/src/ngraph_transformations/convert_to_cpu_specific_opset.hpp

@@ -27,6 +27,7 @@ namespace intel_cpu {
 
 inline void ConvertToCPUSpecificOpset(std::shared_ptr<ngraph::Function> &nGraphFunc) {
     RUN_ON_FUNCTION_SCOPE(ConvertToCPUSpecificOpset);
+
     ngraph::pass::Manager manager;
     manager.register_pass<ConvertMatMulToFC>();
     manager.register_pass<AlignMatMulInputRanks>();