Various minor fixes for plugins

Limit output_dims == W[1] for plugin test
update embLayernorm plugin doc
Remove else after return as they are not required.
Fix FlattenConcat plugin readme

Signed-off-by: Shuyue Lan <shuyuel@nvidia.com>
Signed-off-by: Rajeev Rao <rajeevrao@nvidia.com>

plugin/bertQKVToContextPlugin/qkvToContextPlugin.cpp

@@ -234,7 +234,7 @@ bool QKVToContextPluginDynamic::supportsFormatCombination(
         if (inMask->dims.d[1] != -1 && inMask->dims.d[1] != packedSize)
         {
             gLogError << "CustomEmbLayerNormPluginDynamic returned mask with pack size " << inMask->dims.d[1]
-                      << ", but " << kQKV_TO_CONTEXT_PLUGIN_NAME << " expects mask pack size " << packedSize
+                      << ", but " << QKV_TO_CONTEXT_PLUGIN_NAME << " expects mask pack size " << packedSize
                       << std::endl;
             return false;
         }

plugin/cropAndResizePlugin/cropAndResizePlugin.cpp

@@ -275,10 +275,8 @@ bool CropAndResizePlugin::supportsFormat(DataType type, PluginFormat format) con
     {
         return true;
     }
-    else
-    {
-        return false;
-    }
+
+    return false;
 }
 
 bool CropAndResizeDynamicPlugin::supportsFormatCombination(

plugin/efficientNMSPlugin/efficientNMSPlugin.cpp

@@ -61,11 +61,9 @@ int EfficientNMSPlugin::getNbOutputs() const noexcept
         // ONNX NonMaxSuppression Compatibility
         return 1;
     }
-    else
-    {
-        // Standard Plugin Implementation
-        return 4;
-    }
+
+    // Standard Plugin Implementation
+    return 4;
 }
 
 int EfficientNMSPlugin::initialize() noexcept
@@ -135,16 +133,14 @@ nvinfer1::DataType EfficientNMSPlugin::getOutputDataType(
         // ONNX NMS uses an integer output
         return nvinfer1::DataType::kINT32;
     }
-    else
+
+    // On standard NMS, num_detections and detection_classes use integer outputs
+    if (index == 0 || index == 3)
     {
-        // On standard NMS, num_detections and detection_classes use integer outputs
-        if (index == 0 || index == 3)
-        {
-            return nvinfer1::DataType::kINT32;
-        }
-        // All others should use the same datatype as the input
-        return inputTypes[0];
+        return nvinfer1::DataType::kINT32;
     }
+    // All others should use the same datatype as the input
+    return inputTypes[0];
 }
 
 IPluginV2DynamicExt* EfficientNMSPlugin::clone() const noexcept
@@ -256,14 +252,13 @@ bool EfficientNMSPlugin::supportsFormatCombination(
         return (inOut[pos].type == DataType::kHALF || inOut[pos].type == DataType::kFLOAT)
             && (inOut[0].type == inOut[pos].type);
     }
-    else
+
+    PLUGIN_ASSERT(nbInputs == 2 || nbInputs == 3);
+    PLUGIN_ASSERT(nbOutputs == 4);
+    if (nbInputs == 2)
     {
-        PLUGIN_ASSERT(nbInputs == 2 || nbInputs == 3);
-        PLUGIN_ASSERT(nbOutputs == 4);
-        if (nbInputs == 2)
-        {
-            PLUGIN_ASSERT(0 <= pos && pos <= 5);
-        }
+        PLUGIN_ASSERT(0 <= pos && pos <= 5);
+    }
         if (nbInputs == 3)
         {
             PLUGIN_ASSERT(0 <= pos && pos <= 6);
@@ -279,7 +274,6 @@ bool EfficientNMSPlugin::supportsFormatCombination(
         // all other inputs/outputs: fp32 or fp16
         return (inOut[pos].type == DataType::kHALF || inOut[pos].type == DataType::kFLOAT)
             && (inOut[0].type == inOut[pos].type);
-    }
 }
 
 void EfficientNMSPlugin::configurePlugin(
@@ -387,21 +381,19 @@ int EfficientNMSPlugin::enqueue(const PluginTensorDesc* inputDesc, const PluginT
             return EfficientNMSInference(mParam, boxesInput, scoresInput, nullptr, nullptr, nullptr, nullptr, nullptr,
                 nmsIndicesOutput, workspace, stream);
         }
-        else
-        {
-            // Standard NMS Operation
-            const void* const boxesInput = inputs[0];
-            const void* const scoresInput = inputs[1];
-            const void* const anchorsInput = mParam.boxDecoder ? inputs[2] : nullptr;
 
-            void* numDetectionsOutput = outputs[0];
-            void* nmsBoxesOutput = outputs[1];
-            void* nmsScoresOutput = outputs[2];
-            void* nmsClassesOutput = outputs[3];
+        // Standard NMS Operation
+        void const* const boxesInput = inputs[0];
+        void const* const scoresInput = inputs[1];
+        void const* const anchorsInput = mParam.boxDecoder ? inputs[2] : nullptr;
 
-            return EfficientNMSInference(mParam, boxesInput, scoresInput, anchorsInput, numDetectionsOutput,
-                nmsBoxesOutput, nmsScoresOutput, nmsClassesOutput, nullptr, workspace, stream);
-        }
+        void* numDetectionsOutput = outputs[0];
+        void* nmsBoxesOutput = outputs[1];
+        void* nmsScoresOutput = outputs[2];
+        void* nmsClassesOutput = outputs[3];
+
+        return EfficientNMSInference(mParam, boxesInput, scoresInput, anchorsInput, numDetectionsOutput, nmsBoxesOutput,
+            nmsScoresOutput, nmsClassesOutput, nullptr, workspace, stream);
     }
     catch (const std::exception& e)
     {

plugin/embLayerNormPlugin/README.md

@@ -50,7 +50,7 @@ The final output embedding is the sum of embeddings for the token, the segment a
 
 
 `maskIdx`
-embedded_input is an `int32` tensor with shape `[B,]` where `B` is batch size.
+embedded_input is an `int32` tensor with shape `[B, packSize]` where `B` is batch size, `packSize` is the packed mask size that depends on the sequence length.
 The maskIdx is a more compact representation of the input mask, consisting of the number of valid elements, assuming that the original mask was contiguous.
 
 
@@ -88,10 +88,10 @@ documentation.
 
 ## Changelog
 
-October 2020  
+October 2020
 Add V2 plugin that supports variable sequence length.
 
-November 2019  
+November 2019
 This is the first release of this `README.md` file.
 
 

plugin/fcPlugin/CustomFCPluginDynamic_PluginConfig.yaml

@@ -46,8 +46,7 @@ versions:
         max: "=pinf, =pinf"
     attribute_options:
       out_dims:
-        min: "=1"
-        max: "=pinf"
+        from_shape: "W_1"
       type_id:
         - 0
         - 1

plugin/flattenConcat/README.md

@@ -149,7 +149,7 @@ The following parameters were used to create `FlattenConcat` instance:
 
 | Type             | Parameter                      | Description
 |------------------|--------------------------------|--------------------------------------------------------
-|`int`             |`concatAxis`                    |The dimension along which to concatenate. Currently only `concatAxis = 1` is supported.
+|`int`             |`axis`                          |The dimension along which to concatenate. Currently only `axis = 1` is supported.
 |`bool`            |`ignoreBatch`                   |Whether to ignore batch or not. Currently only `ignoreBatch = false` is supported.
 
 
@@ -174,4 +174,4 @@ This is the first release of this `README.md` file.
 
 ## Known issues
 
-There are no known issues in this plugin.
+There are no known issues in this plugin.

plugin/multiscaleDeformableAttnPlugin/multiscaleDeformableAttnPlugin.cpp

@@ -80,16 +80,10 @@ bool MultiscaleDeformableAttnPlugin::supportsFormatCombination(
         {
             return (inOut[pos].type == nvinfer1::DataType::kINT32);
         }
-        else
-        {
-            return ((inOut[pos].type == inOut[0].type) &&
-                  ((inOut[pos].type == nvinfer1::DataType::kFLOAT) || (inOut[pos].type == nvinfer1::DataType::kHALF)));
-        }
-    }
-    else
-    {
-        return false;
+        return ((inOut[pos].type == inOut[0].type)
+            && ((inOut[pos].type == nvinfer1::DataType::kFLOAT) || (inOut[pos].type == nvinfer1::DataType::kHALF)));
     }
+    return false;
 }
 
 void MultiscaleDeformableAttnPlugin::configurePlugin(nvinfer1::DynamicPluginTensorDesc const* inputs, int32_t nbInputs,

plugin/nvFasterRCNN/nvFasterRCNNPlugin.cpp

@@ -159,10 +159,8 @@ Dims RPROIPlugin::getOutputDimensions(int index, const Dims* inputs, int nbInput
         return Dims3(1, params.nmsMaxOut, 4);
     }
     // Feature map of each ROI after ROI Pooling
-    else // pool5
-    {
-        return Dims4(params.nmsMaxOut, inputs[2].d[0], params.poolingH, params.poolingW);
-    }
+    // pool5
+    return Dims4(params.nmsMaxOut, inputs[2].d[0], params.poolingH, params.poolingW);
 }
 
 size_t RPROIPlugin::getWorkspaceSize(int maxBatchSize) const noexcept

plugin/pillarScatterPlugin/pillarScatter.cpp

@@ -144,48 +144,26 @@ int PillarScatterPlugin::enqueue(const nvinfer1::PluginTensorDesc* inputDesc,
 
         int status = -1;
 
-        if(inputType == nvinfer1::DataType::kHALF){
-            auto pillar_features_data = static_cast<const half *>(inputs[0]);
-            auto spatial_feature_data = static_cast<half *>(outputs[0]);
-            cudaMemsetAsync(spatial_feature_data, 0, batchSize*numFeatures*featureY*featureX * sizeof(half), stream);
-            status = pillarScatterKernelLaunch<half>(
-                batchSize,
-                maxPillarNum,
-                numFeatures,
-                pillar_features_data,
-                coords_data,
-                params_data,
-                featureX,
-                featureY,
-                spatial_feature_data,
-                stream
-                );
-            PLUGIN_ASSERT(status == STATUS_SUCCESS);
-            return status;
+        if (inputType == nvinfer1::DataType::kHALF)
+        {
+            auto pillar_features_data = static_cast<const half*>(inputs[0]);
+            auto spatial_feature_data = static_cast<half*>(outputs[0]);
+            cudaMemsetAsync(
+                spatial_feature_data, 0, batchSize * numFeatures * featureY * featureX * sizeof(half), stream);
+            status = pillarScatterKernelLaunch<half>(batchSize, maxPillarNum, numFeatures, pillar_features_data,
+                coords_data, params_data, featureX, featureY, spatial_feature_data, stream);
         }
-        else if(inputType == nvinfer1::DataType::kFLOAT){
-            auto pillar_features_data = static_cast<const float *>(inputs[0]);
-            auto spatial_feature_data = static_cast<float *>(outputs[0]);
-            cudaMemsetAsync(spatial_feature_data, 0, batchSize*numFeatures*featureY*featureX * sizeof(float), stream);
-            status = pillarScatterKernelLaunch<float>(
-                batchSize,
-                maxPillarNum,
-                numFeatures,
-                pillar_features_data,
-                coords_data,
-                params_data,
-                featureX,
-                featureY,
-                spatial_feature_data,
-                stream
-                );
-            PLUGIN_ASSERT(status == STATUS_SUCCESS);
-            return status;
-        }
-        else{
-            PLUGIN_ASSERT(status == STATUS_SUCCESS);
-            return status;
+        else if (inputType == nvinfer1::DataType::kFLOAT)
+        {
+            auto const* pillar_features_data = static_cast<float const*>(inputs[0]);
+            auto* spatial_feature_data = static_cast<float*>(outputs[0]);
+            cudaMemsetAsync(
+                spatial_feature_data, 0, batchSize * numFeatures * featureY * featureX * sizeof(float), stream);
+            status = pillarScatterKernelLaunch<float>(batchSize, maxPillarNum, numFeatures, pillar_features_data,
+                coords_data, params_data, featureX, featureY, spatial_feature_data, stream);
         }
+        PLUGIN_ASSERT(status == STATUS_SUCCESS);
+        return status;
     }
     catch (const std::exception& e)
     {

plugin/proposalPlugin/proposalPlugin.cpp

@@ -412,10 +412,8 @@ bool ProposalPlugin::supportsFormat(DataType type, PluginFormat format) const no
     {
         return true;
     }
-    else
-    {
-        return false;
-    }
+
+    return false;
 }
 
 bool ProposalDynamicPlugin::supportsFormatCombination(