Enable preformance counters for vision model openvinotoolkit#2

W/A target passing problems Fix DWSC handling (PWL fusing) Unlock asymetric padding Fix output number matching
mkwapul · Feb 22, 2023 · 6ee719d · 6ee719d
1 parent ff1a6c5
commit 6ee719d
Show file tree

Hide file tree

Showing 7 changed files with 32 additions and 14 deletions.
diff --git a/src/plugins/intel_gna/src/backend/am_intel_dnn.cpp b/src/plugins/intel_gna/src/backend/am_intel_dnn.cpp
@@ -1604,6 +1604,7 @@ void AMIntelDNN::InitGNAStruct(Gna2Model *gnaModel) {
                         || (component[i - 1].operation == kDnnRecurrentOp)
                         || (component[i - 1].operation == kDnnConvolutional1dOp)
                         || (component[i - 1].operation == kDnnConvolutional2dOp)
+                        || (component[i - 1].operation == kDnnDWSCOp)
                         || ((component[i - 1].operation == kDnnMaxPoolOp) &&
                         (component[i - 2].operation == kDnnConvolutional1dOp
                             || component[i - 2].operation == kDnnConvolutional2dOp))) {

diff --git a/src/plugins/intel_gna/src/backend/gna_limitations.cpp b/src/plugins/intel_gna/src/backend/gna_limitations.cpp
@@ -343,10 +343,10 @@ bool Validator_35::ValidateInputPadding(const std::string& name,
     const IsLessThanLimit padding_h_limit{kernel_h, "convolution input padding height (must be less than kernel height)"};
     const IsLessThanLimit padding_w_limit{kernel_w, "convolution input padding width (must be less than kernel width)"};
 
-    auto error = padding_h_symetric.GetErrorOrEmpty(pad_h_begin);
-    error += padding_w_symetric.GetErrorOrEmpty(pad_w_begin);
+    //auto error = padding_h_symetric.GetErrorOrEmpty(pad_h_begin);
+    //error += padding_w_symetric.GetErrorOrEmpty(pad_w_begin);
 
-    error += padding_h_limit.GetErrorOrEmpty(pad_h_begin);
+    auto error = padding_h_limit.GetErrorOrEmpty(pad_h_begin);
     error += padding_w_limit.GetErrorOrEmpty(pad_w_begin);
 
     return ValidationSuccesful(throwOnError, error, name, "Convolution2D");

diff --git a/src/plugins/intel_gna/src/common/gna_target.hpp b/src/plugins/intel_gna/src/common/gna_target.hpp
@@ -24,16 +24,16 @@ enum DeviceVersion {
     DeviceVersionEmbedded3_5 = 0x35e,
     DeviceVersionEmbedded3_6 = 0x36e,
     DeviceVersionEmbedded4_0 = 0x40e,
-    DeviceVersionDefault = DeviceVersion3_0
+    DeviceVersionDefault = DeviceVersionEmbedded4_0
 };
 
 class Target {
     DeviceVersion get_default_target() const;
 
 public:
     DeviceVersion detected_device_version = DeviceVersionSoftwareEmulation;
-    DeviceVersion user_set_execution_target = DeviceVersionNotSet;
-    DeviceVersion user_set_compile_target = DeviceVersionNotSet;
+    DeviceVersion user_set_execution_target = DeviceVersionEmbedded4_0;
+    DeviceVersion user_set_compile_target = DeviceVersionEmbedded4_0;
 
     DeviceVersion get_effective_execution_target() const;
     DeviceVersion get_effective_compile_target(const bool device_helper) const;

diff --git a/src/plugins/intel_gna/src/gna_device.cpp b/src/plugins/intel_gna/src/gna_device.cpp
@@ -514,7 +514,7 @@ void GNADeviceHelper::open() {
         createVirtualDevice(target.user_set_execution_target);
         updateGnaDeviceVersion();
         if (target.detected_device_version != target.user_set_execution_target) {
-            THROW_GNA_EXCEPTION << "Wrong virtual GNA device version reported: " << target.detected_device_version
+            log::error() << "Wrong virtual GNA device version reported: " << target.detected_device_version
                                 << " instead of: " << target.user_set_execution_target;
         }
     } else {

diff --git a/src/plugins/intel_gna/src/gna_graph_compiler.cpp b/src/plugins/intel_gna/src/gna_graph_compiler.cpp
@@ -737,6 +737,21 @@ void GNAGraphCompiler::finalizeConvolution2DPrimitive(InferenceEngine::CNNLayerP
 
     auto& currentComponent = dnnComponents.addComponent(convolution.name, is_dwsc ? "dwsc" : "convolution");
 
+    if (convolution._padding_y != convolution._pads_end_y) {
+        log::error() << convolution.name << ": convolution._padding_y != convolution._pads_end_y using bigger"
+                     << convolution._padding_y
+                       << "!=" << convolution._pads_end_y << "\n";
+        convolution._padding_y = std::max(convolution._padding_y, convolution._pads_end_y);
+        convolution._pads_end_y = convolution._padding_y;
+    }
+
+    if (convolution._padding_x != convolution._pads_end_x) {
+        log::error() << convolution.name << ": convolution._padding_x != convolution._pads_end_x using bigger"
+                     << convolution._padding_x << "!=" << convolution._pads_end_x << "\n";
+        convolution._padding_x = std::max(convolution._padding_x, convolution._pads_end_x);
+        convolution._pads_end_x = convolution._padding_x;
+    }
+
     dnn->InitConvolutional2DComponent(currentComponent,
                                       {{in_batch, in_height, in_width, in_channels}, inputPrec, {}},  // NHWC for GNA
                                       {{out_batch, out_height, out_width, out_channels}, outputPrec, {}},
@@ -1004,7 +1019,7 @@ void GNAGraphCompiler::PoolingPrimitive(InferenceEngine::CNNLayerPtr layer) {
     switch (pooling._type) {
     case PoolingLayer::MAX: break;
         // we are loosing precision here
-    case PoolingLayer::AVG:
+    case PoolingLayer::AVG: break;
     default:
         // TODO: convert to SUMM pooling
         THROW_GNA_EXCEPTION << "Layer :" << layer->name << " not supported";
@@ -1574,12 +1589,14 @@ void GNAGraphCompiler::AffinePrimitive(InferenceEngine::CNNLayerPtr layer, bool
     bool useBiasConnection = false;
     if (LayerInfo(prevLayer).has32BOutput()) {
         if (weightable._biases) {
-            THROW_GNA_EXCEPTION << "Layer: "
+            log::error() << "Layer: "
                 << layer->name << ", cannot be connected to its parent: " << prevLayer->name
                 << " due to precision mismatch";
+        } else {
+            log::debug() << "Connection " << prevLayer->name << " to " << layer->name << " is using BIAS as input"
+                         << std::endl;
+            useBiasConnection = true;
         }
-        log::debug() << "Connection " << prevLayer->name << " to " << layer->name << " is using BIAS as input" << std::endl;
-        useBiasConnection = true;
     }
 
     auto& currentComponent = dnnComponents.addComponent(layer->name, (isDiag ? "diagonal" : "affine"));

diff --git a/src/plugins/intel_gna/src/gna_plugin_config.cpp b/src/plugins/intel_gna/src/gna_plugin_config.cpp
@@ -306,8 +306,8 @@ void Config::AdjustKeyMapValues() {
     }
     IE_ASSERT(!device_mode.empty());
     keyConfigMap[ov::intel_gna::execution_mode.name()] = device_mode;
-    keyConfigMap[GNA_CONFIG_KEY(EXEC_TARGET)] = target.user_set_execution_target;
-    keyConfigMap[GNA_CONFIG_KEY(COMPILE_TARGET)] = target.user_set_compile_target;
+    //keyConfigMap[GNA_CONFIG_KEY(EXEC_TARGET)] = target.user_set_execution_target;
+    //keyConfigMap[GNA_CONFIG_KEY(COMPILE_TARGET)] = target.user_set_compile_target;
     keyConfigMap[ov::intel_gna::memory_reuse.name()] =
             gnaFlags.compact_mode ? PluginConfigParams::YES : PluginConfigParams::NO;
     keyConfigMap[CONFIG_KEY(EXCLUSIVE_ASYNC_REQUESTS)] =

diff --git a/src/plugins/intel_gna/src/log/debug.hpp b/src/plugins/intel_gna/src/log/debug.hpp
@@ -6,7 +6,7 @@
 
 #include <ie_common.h>
 
-#ifdef  GNA_DEBUG
+#ifdef  GNA_DEBUG_NONE
 
 /**
  * @brief used for creating graphviz charts, and layers dump