[src] Enable an option to use the GPU for feature extraction in GPU decoding (#3420)

This is turned on by using the option
--gpu-feature-extract=true.  By default this is on.  We provie the
option to turn it off because in situations where CPU resources
are unconfined you can get slightly higher performance with CPU
feature extraction but in most cases GPU feature extraction is faster
and has more stable performance.  In addition a user may wish to
turn it off to support models where feature extraction is currently
incomplete (e.g. FBANK, PLP, PITCH, etc).  We will add those
features in the future but for now a user wanted to decode those
models should place feature extraction on the host.

Author: luitjens

src/cudadecoder/Makefile

@@ -23,7 +23,8 @@ LIBNAME = kaldi-cudadecoder
 ADDLIBS = ../cudamatrix/kaldi-cudamatrix.a ../base/kaldi-base.a ../matrix/kaldi-matrix.a \
           ../lat/kaldi-lat.a ../util/kaldi-util.a ../matrix/kaldi-matrix.a ../gmm/kaldi-gmm.a \
           ../fstext/kaldi-fstext.a ../hmm/kaldi-hmm.a ../gmm/kaldi-gmm.a ../transform/kaldi-transform.a \
-          ../tree/kaldi-tree.a ../online2/kaldi-online2.a ../nnet3/kaldi-nnet3.a
+          ../tree/kaldi-tree.a ../online2/kaldi-online2.a ../nnet3/kaldi-nnet3.a \
+					../cudafeat/kaldi-cudafeat.a
 
 # Implicit rule for kernel compilation
 %.o : %.cu

src/cudadecoder/batched-threaded-nnet3-cuda-pipeline.cc

@@ -248,9 +248,15 @@ void BatchedThreadedNnet3CudaPipeline::OpenDecodeHandle(
   task->callback = std::move(callback);
   task->Init(key, wave_data);
 
-  work_pool_->enqueue(THREAD_POOL_LOW_PRIORITY,
-                      &BatchedThreadedNnet3CudaPipeline::ComputeOneFeature,
-                      this, task);
+  if (config_.gpu_feature_extract) {
+    //Feature extraction done on device
+    AddTaskToPendingTaskQueue(task);
+  } else {
+    //Feature extraction done on host thread
+    work_pool_->enqueue(THREAD_POOL_LOW_PRIORITY,
+                        &BatchedThreadedNnet3CudaPipeline::ComputeOneFeatureCPU,
+                        this, task);
+  }
 }
 
 void BatchedThreadedNnet3CudaPipeline::OpenDecodeHandle(
@@ -262,9 +268,15 @@ void BatchedThreadedNnet3CudaPipeline::OpenDecodeHandle(
   task->callback = std::move(callback);
   task->Init(key, wave_data, sample_rate);
 
-  work_pool_->enqueue(THREAD_POOL_LOW_PRIORITY,
-                      &BatchedThreadedNnet3CudaPipeline::ComputeOneFeature,
-                      this, task);
+  if (config_.gpu_feature_extract) {
+    //Feature extraction done on device
+    AddTaskToPendingTaskQueue(task);
+  } else {
+    //Feature extraction done on host thread
+    work_pool_->enqueue(THREAD_POOL_LOW_PRIORITY,
+                        &BatchedThreadedNnet3CudaPipeline::ComputeOneFeatureCPU,
+                        this, task);
+  }
 }
 
 bool BatchedThreadedNnet3CudaPipeline::GetRawLattice(const std::string &key,
@@ -410,6 +422,8 @@ void BatchedThreadedNnet3CudaPipeline::AquireAdditionalTasks(
 void BatchedThreadedNnet3CudaPipeline::ComputeBatchNnet(
     nnet3::NnetBatchComputer &computer, int32 first,
     std::vector<TaskState *> &tasks) {
+  nvtxRangePushA("ComputeBatchNnet");
+
   bool output_to_cpu = false;
   int32 online_ivector_period = 0;
   int max_pending_minibatches =
@@ -421,17 +435,32 @@ void BatchedThreadedNnet3CudaPipeline::ComputeBatchNnet(
   // for all new batches enqueue up nnet work.
   for (int i = first; i < tasks.size(); i++) {
     TaskState &task = *tasks[i];
-    Vector<BaseFloat> &ivector_features = task.task_data->ivector_features;
-    Matrix<BaseFloat> &input_features = task.task_data->input_features;
+    std::shared_ptr<TaskData> &task_data = task.task_data;
     std::vector<nnet3::NnetInferenceTask> &ntasks = nnet_tasks[i];
+    
+    if (config_.gpu_feature_extract) {
+      CuVector<BaseFloat> &ivector_features = task_data->ivector_features;
+      CuMatrix<BaseFloat> &input_features = task_data->input_features;
+
+      CuVector<BaseFloat> *ifeat = NULL;
+      if (ivector_features.Dim() > 0) {
+        ifeat = &ivector_features;
+      }
+      // create task list
+      computer.SplitUtteranceIntoTasks(output_to_cpu, input_features, ifeat, 
+                                       NULL, online_ivector_period, &ntasks);
+    } else {
+			Vector<BaseFloat> &ivector_features = task_data->ivector_features_cpu;
+      Matrix<BaseFloat> &input_features = task_data->input_features_cpu;
 
-    Vector<BaseFloat> *ifeat = NULL;
-    if (ivector_features.Dim() > 0) {
-      ifeat = &ivector_features;
+      Vector<BaseFloat> *ifeat = NULL;
+      if (ivector_features.Dim() > 0) {
+        ifeat = &ivector_features;
+      }
+      // create task list
+      computer.SplitUtteranceIntoTasks(output_to_cpu, input_features, ifeat, 
+																			 NULL, online_ivector_period, &ntasks);
     }
-    // create task list
-    computer.SplitUtteranceIntoTasks(output_to_cpu, input_features, ifeat, NULL,
-                                     online_ivector_period, &ntasks);
 
     // Add tasks to computer
     for (size_t j = 0; j < ntasks.size(); j++) {
@@ -448,33 +477,37 @@ void BatchedThreadedNnet3CudaPipeline::ComputeBatchNnet(
   // Extract Posteriors
   for (int i = first; i < tasks.size(); i++) {
     TaskState &task = *tasks[i];
-    CuMatrix<BaseFloat> &posteriors = task.task_data->posteriors;
+    std::shared_ptr<TaskData> &task_data = task.task_data;
+    CuMatrix<BaseFloat> &posteriors = task_data->posteriors;
     MergeTaskOutput(nnet_tasks[i], &posteriors);
 
     // nnet output is no longer necessary as we have copied the output out
     nnet_tasks[i].resize(0);
 
     // featurs are no longer needed so free memory
-    task.task_data->ivector_features.Resize(0);
-    task.task_data->input_features.Resize(0, 0);
+    task_data->ivector_features.Resize(0);
+    task_data->input_features.Resize(0, 0);
   }
+
+  nvtxRangePop();
 }
 
 // Computes Features for a single decode instance.
-void BatchedThreadedNnet3CudaPipeline::ComputeOneFeature(TaskState *task_) {
-  nvtxRangePushA("ComputeOneFeature");
+void BatchedThreadedNnet3CudaPipeline::ComputeOneFeatureCPU(TaskState *task_) {
+  nvtxRangePushA("ComputeOneFeatureCPU");
   TaskState &task = *task_;
-  Vector<BaseFloat> &ivector_features = task.task_data->ivector_features;
-  Matrix<BaseFloat> &input_features = task.task_data->input_features;
+  std::shared_ptr<TaskData> &task_data = task.task_data;
+  Vector<BaseFloat> &ivector_features = task_data->ivector_features_cpu;
+  Matrix<BaseFloat> &input_features = task_data->input_features_cpu;
 
   // create decoding state
   OnlineNnet2FeaturePipeline feature(*feature_info_);
 
   // Accept waveforms
   feature.AcceptWaveform(
-      task.task_data->sample_frequency,
-      SubVector<BaseFloat>(*task.task_data->wave_samples, 0,
-                           task.task_data->wave_samples->Dim()));
+      task_data->sample_frequency,
+      SubVector<BaseFloat>(*task_data->wave_samples, 0,
+                           task_data->wave_samples->Dim()));
   feature.InputFinished();
   // All frames should be ready here
   int32 numFrames = feature.NumFramesReady();
@@ -487,7 +520,8 @@ void BatchedThreadedNnet3CudaPipeline::ComputeOneFeature(TaskState *task_) {
 
   std::vector<int> frames(numFrames);
   // create list of frames
-  for (int j = 0; j < numFrames; j++) frames[j] = j;
+  for (int j = 0; j < numFrames; j++) 
+    frames[j] = j;
 
   // Copy Features
   input_features.Resize(numFrames, input_dim);
@@ -501,18 +535,106 @@ void BatchedThreadedNnet3CudaPipeline::ComputeOneFeature(TaskState *task_) {
     // Copy Features
     feature.IvectorFeature()->GetFrame(numFrames - 1, &ivector_features);
   }
-  nvtxRangePop();
 
   AddTaskToPendingTaskQueue(task_);
+
+  nvtxRangePop();
 }
 
+// Computes features across the tasks[first,tasks.size()
+void BatchedThreadedNnet3CudaPipeline::ComputeBatchFeatures(
+    int32 first, std::vector<TaskState *> &tasks, 
+    OnlineCudaFeaturePipeline &feature_pipeline) {
+  KALDI_ASSERT(config_.gpu_feature_extract==true);
+  nvtxRangePushA("CopyBatchWaves");
+  // below we will pack waves into a single buffer for efficient transfer across device
+  
+  // first count the total number of elements and create a single large vector
+  int count=0;
+  for (int i = first; i < tasks.size(); i++) {
+    count+=tasks[i]->task_data->wave_samples->Dim();
+  }
+
+  // creating a thread local vector of pinned memory.
+  // wave data will be stagged through this memory to get 
+  // more efficient non-blocking transfers to the device.
+  thread_local Vector<BaseFloat> pinned_vector;
+  
+  if (pinned_vector.Dim() < count ) {
+    if ( pinned_vector.Dim()!=0) {
+      cudaHostUnregister(pinned_vector.Data());
+    }
+    // allocated array 2x size
+    pinned_vector.Resize(count*2,kUndefined);
+    cudaHostRegister(pinned_vector.Data(), pinned_vector.Dim()*sizeof(BaseFloat),0);
+  }
+
+  // We will launch a thread for each task in order to get better host memory bandwidth
+  std::vector<std::future<void> > futures;  //for syncing
+
+  //vector copy function for threading below.
+  auto copy_vec = [](SubVector<BaseFloat> &dst, const SubVector<BaseFloat> &src) { 
+    nvtxRangePushA("CopyVec");
+    dst.CopyFromVec(src); 
+    nvtxRangePop();
+  };
+
+  // next launch threads to copy all waves for each task in parallel
+  count=0;
+  for (int i = first; i < tasks.size(); i++) {
+    std::shared_ptr<TaskData> &task_data = tasks[i]->task_data;
+    SubVector<BaseFloat> wave(pinned_vector,count,task_data->wave_samples->Dim());
+    count+=task_data->wave_samples->Dim();
+    futures.push_back(
+      work_pool_->enqueue(copy_vec, wave, *(task_data->wave_samples))
+      );
+  }
+
+  // wait for waves to be copied into place
+  for (int i = 0; i < futures.size(); i++) {
+    futures[i].get();
+  }
+
+  CuVector<BaseFloat> cu_waves(count, kUndefined);
+  // copy memory down asynchronously.  Vector copy functions are synchronous so we do it manually.
+  // It is important for this to happen asynchrously to help hide launch latency of smaller kernels
+  // that come in the future.
+  cudaMemcpyAsync(cu_waves.Data(), pinned_vector.Data(), cu_waves.Dim()*sizeof(BaseFloat), 
+      cudaMemcpyHostToDevice, cudaStreamPerThread);
+  nvtxRangePop();
+  
+  nvtxRangePushA("ComputeBatchFeatures");
+  // extract features for each wave
+  count=0;
+  for (int i = first; i < tasks.size(); i++) {
+    TaskState &task = *tasks[i];
+    std::shared_ptr<TaskData> &task_data = task.task_data;
+  
+    CuSubVector<BaseFloat> cu_wave(cu_waves,count,task_data->wave_samples->Dim());
+    count+=task_data->wave_samples->Dim();
+    feature_pipeline.ComputeFeatures(cu_wave, task_data->sample_frequency, 
+        &task_data->input_features, &task_data->ivector_features);
+  
+    int32 numFrames = task_data->input_features.NumRows();
+
+    if (numFrames == 0) {
+      //Make this a warning for now.  Need to check how this is handled
+      KALDI_WARN << "Warning empty audio file";
+    }
+  }
+  nvtxRangePop();
+}
+
+
+
 // Allocates decodables for tasks in the range of tasks[first,tasks.size())
 void BatchedThreadedNnet3CudaPipeline::AllocateDecodables(
     int32 first, std::vector<TaskState *> &tasks,
     std::vector<CudaDecodableInterface *> &decodables) {
   // Create mapped decodable here
   for (int i = first; i < tasks.size(); i++) {
-    CuMatrix<BaseFloat> &posteriors = tasks[i]->task_data->posteriors;
+    std::shared_ptr<TaskData> &task_data = tasks[i]->task_data;
+    CuMatrix<BaseFloat> &posteriors = task_data->posteriors;
     decodables.push_back(
         new DecodableCuMatrixMapped(*trans_model_, posteriors, 0));
   }
@@ -666,6 +788,8 @@ void BatchedThreadedNnet3CudaPipeline::ExecuteWorker(int threadId) {
   nnet3::NnetBatchComputer computer(config_.compute_opts, am_nnet_->GetNnet(),
                                     am_nnet_->Priors());
 
+  OnlineCudaFeaturePipeline feature_pipeline(config_.feature_opts);
+
   ChannelState channel_state;
 
   std::vector<TaskState *> tasks;  // The state for each decode
@@ -713,10 +837,12 @@ void BatchedThreadedNnet3CudaPipeline::ExecuteWorker(int threadId) {
 
         // New tasks are now in the in tasks[start,tasks.size())
         if (start != tasks.size()) {  // if there are new tasks
+					if (config_.gpu_feature_extract)
+            ComputeBatchFeatures(start, tasks, feature_pipeline);
           ComputeBatchNnet(computer, start, tasks);
           AllocateDecodables(start, tasks, decodables);
         }
-      }  // end if(tasks_front_!=tasks_back_)
+      }  // end if (tasks_front_!=tasks_back_)
 
       // check if there is no active work on this thread.
       // This can happen if another thread was assigned the work.

src/cudadecoder/batched-threaded-nnet3-cuda-pipeline.h

@@ -27,6 +27,7 @@
 #include "lat/determinize-lattice-pruned.h"
 #include "nnet3/nnet-batch-compute.h"
 #include "online2/online-nnet2-feature-pipeline.h"
+#include "cudafeat/online-cuda-feature-pipeline.h"
 #include "thread-pool.h"
 
 // If num_channels sets to automatic,
@@ -53,7 +54,8 @@ struct BatchedThreadedNnet3CudaPipelineConfig {
         num_worker_threads(20),
         determinize_lattice(true),
         max_pending_tasks(4000),
-        num_decoder_copy_threads(2){};
+        num_decoder_copy_threads(2),
+        gpu_feature_extract(true) {};
   void Register(OptionsItf *po) {
     po->Register("max-batch-size", &max_batch_size,
                  "The maximum batch size to be used by the decoder. "
@@ -88,6 +90,11 @@ struct BatchedThreadedNnet3CudaPipelineConfig {
     po->Register("cuda-decoder-copy-threads", &num_decoder_copy_threads,
                  "Advanced - Number of worker threads used in the decoder for "
                  "the host to host copies.");
+    po->Register("gpu-feature-extract", &gpu_feature_extract,
+                 "Extract features on the GPU.  This reduces CPU overhead "
+                 "leading to better scalability but may reduce overall "
+                 "performance for a single GPU.");
+
     feature_opts.Register(po);
     decoder_opts.Register(po);
     det_opts.Register(po);
@@ -101,6 +108,7 @@ struct BatchedThreadedNnet3CudaPipelineConfig {
   bool determinize_lattice;
   int max_pending_tasks;
   int num_decoder_copy_threads;
+  bool gpu_feature_extract;
 
   void ComputeConfig() {
     if (num_channels == -1)
@@ -203,8 +211,10 @@ class BatchedThreadedNnet3CudaPipeline {
        wave_samples;  // Used as a pointer to either the raw
                       // data or the samples passed
    float sample_frequency;
-   Vector<BaseFloat> ivector_features;
-   Matrix<BaseFloat> input_features;
+   Vector<BaseFloat> ivector_features_cpu;
+   Matrix<BaseFloat> input_features_cpu;
+   CuVector<BaseFloat> ivector_features;
+   CuMatrix<BaseFloat> input_features;
    CuMatrix<BaseFloat> posteriors;
 
    TaskData(const WaveData &wave_data_in)
@@ -289,7 +299,12 @@ class BatchedThreadedNnet3CudaPipeline {
                              std::vector<TaskState *> &tasks);
 
   // Computes Features for a single decode instance.
-  void ComputeOneFeature(TaskState *task);
+  void ComputeOneFeatureCPU(TaskState *task);
+
+  // Computes features across the tasks[first,tasks.size()
+  void ComputeBatchFeatures(int32 first,
+                            std::vector<TaskState *> &tasks,
+                            OnlineCudaFeaturePipeline &feature_pipeline);
 
   // Computes Nnet across the current decode batch
   void ComputeBatchNnet(nnet3::NnetBatchComputer &computer, int32 first,