make sure nvtx3 is used (#4937)

Author: jtrmal

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

@@ -26,7 +26,7 @@
 
 #include "hipify.h"
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include <mutex>

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

@@ -31,7 +31,7 @@
 
 #include "hipify.h"
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include "base/kaldi-utils.h"

src/cudadecoder/batched-threaded-nnet3-cuda-pipeline2.cc

@@ -28,7 +28,7 @@
 
 #include "hipify.h"
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 namespace kaldi {

src/cudadecoder/cuda-decoder.cc

@@ -44,7 +44,7 @@
 #include "hipify.h"
 #else
 #include <cuda_runtime_api.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include "base/kaldi-utils.h"

src/cudadecoder/cuda-fst.cc

@@ -29,7 +29,7 @@
 #include "hipify.h"
 #else
 #include <cuda_runtime_api.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 namespace kaldi {

src/cudadecoderbin/batched-wav-nnet3-cuda-online.cc

@@ -30,7 +30,7 @@
 #else
 #include <cuda.h>
 #include <cuda_profiler_api.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include <algorithm>

src/cudadecoderbin/batched-wav-nnet3-cuda.cc

@@ -24,7 +24,7 @@
 #else
 #include <cuda.h>
 #include <cuda_profiler_api.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 #include <sstream>
 #include "cudadecoder/batched-threaded-nnet3-cuda-pipeline.h"

src/cudadecoderbin/batched-wav-nnet3-cuda2.cc

@@ -27,7 +27,7 @@
 #else
 #include <cuda.h>
 #include <cuda_profiler_api.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include <sstream>

src/cudafeat/feature-online-batched-spectral-cuda-kernels.cu

@@ -25,7 +25,7 @@
 #include "hipify.h"
 #else
 #include <cub/cub.cuh>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include "cudafeat/lane-desc.h"

src/cudafeat/feature-spectral-cuda.cu

@@ -24,7 +24,7 @@
 
 #include "hipify.h"
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #include <cub/cub.cuh>
 #endif
 

src/cudafeat/feature-window-cuda.cu

@@ -22,7 +22,7 @@
 
 #include "hipify.h"
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 #include "matrix/matrix-functions.h"

src/cudafeat/online-batched-feature-pipeline-cuda.cc

@@ -25,7 +25,7 @@
 
 #include "hipify.h"
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 
 namespace kaldi {

src/cudafeat/online-ivector-feature-cuda.cc

@@ -26,7 +26,7 @@
 #endif
 #define CUBLAS_FILL_MODE_LOWER HIPSOLVER_FILL_MODE_LOWER
 #else
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 #endif
 

src/cudafeatbin/compute-online-feats-batched-cuda.cc

@@ -18,7 +18,7 @@
 #if HAVE_CUDA
 #ifndef __IS_HIP_COMPILE__
 #include <cuda_profiler_api.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 #endif
 

src/cudafeatbin/compute-online-feats-cuda.cc

@@ -17,7 +17,7 @@
 
 #if HAVE_CUDA == 1
 #ifndef __IS_HIP_COMPILE__
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 #endif
 #endif
 #include "base/kaldi-common.h"

src/cudamatrix/cu-common.h

@@ -44,7 +44,7 @@
 #include <cuda_runtime_api.h>
 #include <curand.h>
 #include <cusparse.h>
-#include <nvToolsExt.h>
+#include <nvtx3/nvToolsExt.h>
 
 #define GPU_WARP_SIZE 32
 #define GPU_MAX_THREADS_PER_BLOCK 1024

src/makefiles/cuda_64bit.mk

@@ -16,4 +16,4 @@ CUDA_FLAGS = --compiler-options -fPIC --machine 64 -DHAVE_CUDA -DOPENFST_VER=$(O
 CUDA_LDFLAGS += -L$(CUDATKDIR)/lib64/stubs -L$(CUDATKDIR)/lib64 -Wl,-rpath,$(CUDATKDIR)/lib64
 CUDA_LDFLAGS += -L$(CUDATKDIR)/lib/stubs -L$(CUDATKDIR)/lib -Wl,-rpath,$(CUDATKDIR)/lib
 
-CUDA_LDLIBS += -lcuda -lcublas -lcusparse -lcusolver -lcudart -lcurand -lcufft -lnvToolsExt
+CUDA_LDLIBS += -lcuda -lcublas -lcusparse -lcusolver -lcudart -lcurand -lcufft