Jetson/L4T/TRT Customized Example

From eLinux.org
< Jetson‎ | L4T
Jump to: navigation, search

This page collects information to deploy customized models with TensorRT and some common questions for Jetson.

TensorRT Python

OpenCV with ONNX model

Below is an example to deploy TensorRT from an ONNX model with OpenCV images.

Verified environment:

  • JetPack4.5.1 + Xavier
import cv2
import time
import numpy as np
import tensorrt as trt
import pycuda.autoinit
import pycuda.driver as cuda

EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
runtime = trt.Runtime(TRT_LOGGER)

host_inputs  = []
cuda_inputs  = []
host_outputs = []
cuda_outputs = []
bindings = []


def Inference(engine):
    image = cv2.imread("/usr/src/tensorrt/data/resnet50/airliner.ppm")
    image = (2.0 / 255.0) * image.transpose((2, 0, 1)) - 1.0

    np.copyto(host_inputs[0], image.ravel())
    stream = cuda.Stream()
    context = engine.create_execution_context()

    start_time = time.time()
    cuda.memcpy_htod_async(cuda_inputs[0], host_inputs[0], stream)
    context.execute_async(bindings=bindings, stream_handle=stream.handle)
    cuda.memcpy_dtoh_async(host_outputs[0], cuda_outputs[0], stream)
    stream.synchronize()
    print("execute times "+str(time.time()-start_time))

    output = host_outputs[0].reshape(np.concatenate(([1],engine.get_binding_shape(1))))
    print(np.argmax(output))


def PrepareEngine():
    with trt.Builder(TRT_LOGGER) as builder, builder.create_network(EXPLICIT_BATCH) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
        builder.max_workspace_size = 1 << 30
        with open('/usr/src/tensorrt/data/resnet50/ResNet50.onnx', 'rb') as model:
            if not parser.parse(model.read()):
                print ('ERROR: Failed to parse the ONNX file.')
                for error in range(parser.num_errors):
                    print (parser.get_error(error))
        engine = builder.build_cuda_engine(network)

        # create buffer
        for binding in engine:
            size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
            host_mem = cuda.pagelocked_empty(shape=[size],dtype=np.float32)
            cuda_mem = cuda.mem_alloc(host_mem.nbytes)

            bindings.append(int(cuda_mem))
            if engine.binding_is_input(binding):
                host_inputs.append(host_mem)
                cuda_inputs.append(cuda_mem)
            else:
                host_outputs.append(host_mem)
                cuda_outputs.append(cuda_mem)

        return engine


if __name__ == "__main__":
    engine = PrepareEngine()
    Inference(engine)


OpenCV with PLAN model

Below is an example to deploy TensorRT from a TensorRT PLAN model with OpenCV images.

Verified environment:

  • JetPack4.5.1 + Xavier
$ /usr/src/tensorrt/bin/trtexec --onnx=/usr/src/tensorrt/data/resnet50/ResNet50.onnx --saveEngine=trt.plan
import cv2
import time
import numpy as np
import tensorrt as trt
import pycuda.autoinit
import pycuda.driver as cuda

EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
runtime = trt.Runtime(TRT_LOGGER)

host_inputs  = []
cuda_inputs  = []
host_outputs = []
cuda_outputs = []
bindings = []


def Inference(engine):
    image = cv2.imread("/usr/src/tensorrt/data/resnet50/airliner.ppm")
    image = (2.0 / 255.0) * image.transpose((2, 0, 1)) - 1.0

    np.copyto(host_inputs[0], image.ravel())
    stream = cuda.Stream()
    context = engine.create_execution_context()

    start_time = time.time()
    cuda.memcpy_htod_async(cuda_inputs[0], host_inputs[0], stream)
    context.execute_async(bindings=bindings, stream_handle=stream.handle)
    cuda.memcpy_dtoh_async(host_outputs[0], cuda_outputs[0], stream)
    stream.synchronize()
    print("execute times "+str(time.time()-start_time))

    output = host_outputs[0].reshape(np.concatenate(([1],engine.get_binding_shape(1))))
    print(np.argmax(output))


def PrepareEngine():
    runtime = trt.Runtime(TRT_LOGGER)
    with open('./trt.plan', 'rb') as f:
        buf = f.read()
        engine = runtime.deserialize_cuda_engine(buf)

    # create buffer
    for binding in engine:
        size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
        host_mem = cuda.pagelocked_empty(shape=[size],dtype=np.float32)
        cuda_mem = cuda.mem_alloc(host_mem.nbytes)

        bindings.append(int(cuda_mem))
        if engine.binding_is_input(binding):
            host_inputs.append(host_mem)
            cuda_inputs.append(cuda_mem)
        else:
            host_outputs.append(host_mem)
            cuda_outputs.append(cuda_mem)

    return engine


if __name__ == "__main__":
    engine = PrepareEngine()
    Inference(engine)


Multi-threading

Below is an example to run TensorRT with threads.

Verified environment:

  • JetPack4.5.1 + Xavier
$ /usr/src/tensorrt/bin/trtexec --onnx=/usr/src/tensorrt/data/mnist/mnist.onnx --saveEngine=mnist.trt
$ cd /usr/src/tensorrt/data/mnist/
$ sudo pip3 install pillow
$ python3 download_pgms.py
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/my_tensorrt_code.py -O my_tensorrt_code.py
import threading
import time
from my_tensorrt_code import TRTInference, trt

exitFlag = 0

class myThread(threading.Thread):
   def __init__(self, func, args):
      threading.Thread.__init__(self)
      self.func = func
      self.args = args
   def run(self):
      print ("Starting " + self.args[0])
      self.func(*self.args)
      print ("Exiting " + self.args[0])

if __name__ == '__main__':
    # Create new threads
    '''
    format thread:
        - func: function names, function that we wished to use
        - arguments: arguments that will be used for the func's arguments
    '''

    trt_engine_path = 'mnist.trt'

    max_batch_size = 1
    trt_inference_wrapper = TRTInference(trt_engine_path,
        trt_engine_datatype=trt.DataType.FLOAT,
        batch_size=max_batch_size)

    # Get TensorRT SSD model output
    input_img_path = '/usr/src/tensorrt/data/mnist/3.pgm'

    thread1 = myThread(trt_inference_wrapper.infer, [input_img_path])

    # Start new Threads
    thread1.start()
    thread1.join()
    trt_inference_wrapper.destory();
    print ("Exiting Main Thread")


Deepstream

YoloV4 Tiny

Verified environment:

  • JetPack4.5.1 + Xavier

Deepstream can reach 60fps with 4 video stream on Xavier:

$ cd /opt/nvidia/deepstream/deepstream-5.1/sources/objectDetector_Yolo
$ wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/yolov4-tiny/yolov4_tiny.patch
$ git apply yolov4_tiny.patch
$ export CUDA_VER=10.2
$ make -C nvdsinfer_custom_impl_Yolo
$ wget https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg -q --show-progress
$ wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights -q --show-progress
$ wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/yolov4-tiny/deepstream_app_config_yoloV4_tiny.txt
$ wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/yolov4-tiny/config_infer_primary_yoloV4_tiny.txt
$ deepstream-app -c deepstream_app_config_yoloV4_tiny.txt


Custom Parser for SSD-MobileNet Trained by Jetson-inference

Verified environment:

  • JetPack4.5.1 + Xavier
$ cd /opt/nvidia/deepstream/deepstream-5.1/sources/objectDetector_SSD/
$ sudo wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/ssd-jetson_inference/ssd-jetson_inference.patch
$ sudo git apply ssd-jetson_inference.patch
$ sudo CUDA_VER=10.2 make -C nvdsinfer_custom_impl_ssd/

Update config_infer_primary_ssd.txt:

Ex.

diff --git a/config_infer_primary_ssd.txt b/config_infer_primary_ssd.txt
index e5bf468..81c52fd 100644
--- a/config_infer_primary_ssd.txt
+++ b/config_infer_primary_ssd.txt
@@ -62,15 +62,13 @@ gpu-id=0
 net-scale-factor=0.0078431372
 offsets=127.5;127.5;127.5
 model-color-format=0
-model-engine-file=sample_ssd_relu6.uff_b1_gpu0_fp32.engine
-labelfile-path=ssd_coco_labels.txt
-uff-file=sample_ssd_relu6.uff
+model-engine-file=ssd-mobilenet.uff_b1_gpu0_fp16.engine
+uff-file=ssd.uff
 infer-dims=3;300;300
 uff-input-order=0
 uff-input-blob-name=Input
-batch-size=1
-## 0=FP32, 1=INT8, 2=FP16 mode
-network-mode=0
+labelfile-path=labels.txt
+network-mode=2
 num-detected-classes=91
 interval=0
 gie-unique-id=1
$ deepstream-app -c deepstream_app_config_ssd.txt


VPI

VPI with Jetson-utils

Below is an example to use VPI with jetson-utils

Verified environment:

  • JetPack4.6 + XavierNX
import numpy as np
import jetson.utils
import vpi


display = jetson.utils.glDisplay()

camera = jetson.utils.gstCamera(1920, 1280, '0')
camera.Open()

while display.IsOpen():
    frame, width, height = camera.CaptureRGBA(zeroCopy=1)
    input = vpi.asimage(np.uint8(jetson.utils.cudaToNumpy(frame)))
    with vpi.Backend.CUDA:
        output = input.convert(vpi.Format.U8) 
        output = output.box_filter(11, border=vpi.Border.ZERO).convert(vpi.Format.RGB8)
        vpi.clear_cache()

    display.RenderOnce(jetson.utils.cudaFromNumpy(output.cpu()), width, height)
    display.SetTitle("{:s} | {:d}x{:d} | {:.1f} FPS".format("Camera Viewer", width, height, display.GetFPS()))

camera.Close()


VPI with Deepstream

Please find the following link for the example:

https://forums.developer.nvidia.com/t/deepstream-sdk-vpi-on-jetson-tx2/166834/20

VPI with Argus Camera

Please find the following link for the example:

https://forums.developer.nvidia.com/t/how-do-i-get-image-from-cudabayerdemosaic-and-connect-to-vpi/213529/18

VPI with nvivafilter

Please find the following link for the example:

https://forums.developer.nvidia.com/t/using-vpi-in-gstreamer/223334/21


Stress Test for Orin

We describe the testing tools that can stress the Jetson AGX Orin to the full workload.

The expected power consumption with these steps is listed in the below table:

Jetson AGX Orin 64GB Jetson AGX Orin 32GB
Maximum Power 60W 40W

Maximize the Device Performance

$ sudo nvpmodel -m 0
$ sudo jetson_clocks

CPU Stress Test

Using Linux stress tool:

$ sudo apt-get install stress
$ stress --cpu $(nproc)

GPU Stress Test

Running cuBLAS sample with the half data type:

1. Find matrixMulCUBLAS sample under CUDA sample folder

2. Apply the following change the data type from float to half

diff --git a/Makefile b/Makefile
index 9f5db72..c2a5190 100644
--- a/Makefile
+++ b/Makefile
@@ -274,7 +274,7 @@ ALL_LDFLAGS += $(addprefix -Xlinker ,$(LDFLAGS))
 ALL_LDFLAGS += $(addprefix -Xlinker ,$(EXTRA_LDFLAGS))
 
 # Common includes and paths for CUDA
-INCLUDES  := -I../../common/inc
+INCLUDES  := -I$(CUDA_PATH)/samples/common/inc
 LIBRARIES :=
 
 ################################################################################
@@ -354,14 +354,11 @@ matrixMulCUBLAS.o:matrixMulCUBLAS.cpp
 
 matrixMulCUBLAS: matrixMulCUBLAS.o
 	$(EXEC) $(NVCC) $(ALL_LDFLAGS) $(GENCODE_FLAGS) -o $@ $+ $(LIBRARIES)
-	$(EXEC) mkdir -p ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
-	$(EXEC) cp $@ ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)
 
 run: build
 	$(EXEC) ./matrixMulCUBLAS
 
 clean:
 	rm -f matrixMulCUBLAS matrixMulCUBLAS.o
-	rm -rf ../../bin/$(TARGET_ARCH)/$(TARGET_OS)/$(BUILD_TYPE)/matrixMulCUBLAS
 
 clobber: clean
diff --git a/matrixMulCUBLAS.cpp b/matrixMulCUBLAS.cpp
index 132afd4..8e96a43 100644
--- a/matrixMulCUBLAS.cpp
+++ b/matrixMulCUBLAS.cpp
@@ -72,6 +72,52 @@ typedef struct _matrixSize      // Optional Command-line multiplier for matrix s
     unsigned int uiWA, uiHA, uiWB, uiHB, uiWC, uiHC;
 } sMatrixSize;
 
+// Modified from helper_image.h to support half precision arguments
+inline bool sdkCompareL2fe(const __half *reference, const __half *data,
+                           const unsigned int len, const float epsilon) {
+  assert(epsilon >= 0);
+
+  float error = 0;
+  float ref = 0;
+
+  for (unsigned int i = 0; i < len; ++i) {
+    float diff = (float)reference[i] - (float)data[i];
+    error += diff * diff;
+    ref += (float)reference[i] * (float)reference[i];
+  }
+
+  float normRef = sqrtf(ref);
+
+  if (fabs(ref) < 1e-7) {
+#ifdef _DEBUG
+    std::cerr << "ERROR, reference l2-norm is 0\n";
+#endif
+    return false;
+  }
+
+  float normError = sqrtf(error);
+  error = normError / normRef;
+  bool result = error < epsilon;
+#ifdef _DEBUG
+
+  if (!result) {
+    std::cerr << "ERROR, l2-norm error " << error << " is greater than epsilon "
+              << epsilon << "\n";
+  }
+
+#endif
+
+  return result;
+}
+
+inline int idx_n(int row, int col, int height, int width) {
+    return row * width + col;
+}
+
+inline int idx_t(int row, int col, int height, int width) {
+    return col * height + row;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 //! Compute reference data set matrix multiply on CPU
 //! C = A * B
@@ -82,32 +128,34 @@ typedef struct _matrixSize      // Optional Command-line multiplier for matrix s
 //! @param wB         width of matrix B
 ////////////////////////////////////////////////////////////////////////////////
 void
-matrixMulCPU(float *C, const float *A, const float *B, unsigned int hA, unsigned int wA, unsigned int wB)
+matrixMulCPU(__half *C, const __half *A, const __half *B, unsigned int hA, unsigned int wA, unsigned int wB, const bool ta, const bool tb)
 {
+    auto idx_a = (ta) ? idx_t : idx_n;
+    auto idx_b = (tb) ? idx_t : idx_n;
     for (unsigned int i = 0; i < hA; ++i)
         for (unsigned int j = 0; j < wB; ++j)
         {
-            double sum = 0;
+            float sum = 0;
 
             for (unsigned int k = 0; k < wA; ++k)
             {
-                double a = A[i * wA + k];
-                double b = B[k * wB + j];
+                __half a = A[idx_a(i, k, hA, wA)];
+                __half b = B[idx_b(k, j, wA, wB)];
                 sum += a * b;
             }
 
-            C[i * wB + j] = (float)sum;
+            C[i * wB + j] = sum;
         }
 }
 
-// Allocates a matrix with random float entries.
-void randomInit(float *data, int size)
+// Allocates a matrix with random __half entries.
+void randomInit(__half *data, int size)
 {
     for (int i = 0; i < size; ++i)
-        data[i] = rand() / (float)RAND_MAX;
+        data[i] = (__half)(rand() / (float)(RAND_MAX));
 }
 
-void printDiff(float *data1, float *data2, int width, int height, int iListLength, float fListTol)
+void printDiff(__half *data1, __half *data2, int width, int height, int iListLength, float fListTol)
 {
     printf("Listing first %d Differences > %.6f...\n", iListLength, fListTol);
     int i,j,k;
@@ -129,7 +177,7 @@ void printDiff(float *data1, float *data2, int width, int height, int iListLengt
             {
                 if (error_count < iListLength)
                 {
-                    printf("    Loc(%d,%d)\tCPU=%.5f\tGPU=%.5f\tDiff=%.6f\n", i, j, data1[k], data2[k], fDiff);
+                    printf("    Loc(%d,%d)\tCPU=%.5f\tGPU=%.5f\tDiff=%.6f\n", i, j, (float)data1[k], (float)data2[k], fDiff);
                 }
 
                 error_count++;
@@ -170,12 +218,13 @@ void initializeCUDA(int argc, char **argv, int &devID, int &iSizeMultiple, sMatr
 
     int block_size = 32;
 
-    matrix_size.uiWA = 3 * block_size * iSizeMultiple;
-    matrix_size.uiHA = 4 * block_size * iSizeMultiple;
-    matrix_size.uiWB = 2 * block_size * iSizeMultiple;
-    matrix_size.uiHB = 3 * block_size * iSizeMultiple;
-    matrix_size.uiWC = 2 * block_size * iSizeMultiple;
-    matrix_size.uiHC = 4 * block_size * iSizeMultiple;
+    const int N = 2048;
+    matrix_size.uiWA = N;//3 * block_size * iSizeMultiple;
+    matrix_size.uiHA = N;//4 * block_size * iSizeMultiple;
+    matrix_size.uiWB = N;//2 * block_size * iSizeMultiple;
+    matrix_size.uiHB = N;//3 * block_size * iSizeMultiple;
+    matrix_size.uiWC = N;//2 * block_size * iSizeMultiple;
+    matrix_size.uiHC = N;//4 * block_size * iSizeMultiple;
 
     printf("MatrixA(%u,%u), MatrixB(%u,%u), MatrixC(%u,%u)\n",
            matrix_size.uiHA, matrix_size.uiWA,
@@ -194,7 +243,7 @@ void initializeCUDA(int argc, char **argv, int &devID, int &iSizeMultiple, sMatr
 ////////////////////////////////////////////////////////////////////////////////
 //! Run a simple test matrix multiply using CUBLAS
 ////////////////////////////////////////////////////////////////////////////////
-int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
+int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size, bool ta, bool tb)
 {
     cudaDeviceProp deviceProp;
 
@@ -207,11 +256,11 @@ int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
 
     // allocate host memory for matrices A and B
     unsigned int size_A = matrix_size.uiWA * matrix_size.uiHA;
-    unsigned int mem_size_A = sizeof(float) * size_A;
-    float *h_A = (float *)malloc(mem_size_A);
+    unsigned int mem_size_A = sizeof(__half) * size_A;
+    __half *h_A = (__half *)malloc(mem_size_A);
     unsigned int size_B = matrix_size.uiWB * matrix_size.uiHB;
-    unsigned int mem_size_B = sizeof(float) * size_B;
-    float *h_B = (float *)malloc(mem_size_B);
+    unsigned int mem_size_B = sizeof(__half) * size_B;
+    __half *h_B = (__half *)malloc(mem_size_B);
 
     // set seed for rand()
     srand(2006);
@@ -221,13 +270,13 @@ int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
     randomInit(h_B, size_B);
 
     // allocate device memory
-    float *d_A, *d_B, *d_C;
+    __half *d_A, *d_B, *d_C;
     unsigned int size_C = matrix_size.uiWC * matrix_size.uiHC;
-    unsigned int mem_size_C = sizeof(float) * size_C;
+    unsigned int mem_size_C = sizeof(__half) * size_C;
 
     // allocate host memory for the result
-    float *h_C      = (float *) malloc(mem_size_C);
-    float *h_CUBLAS = (float *) malloc(mem_size_C);
+    __half *h_C      = (__half *) malloc(mem_size_C);
+    __half *h_CUBLAS = (__half *) malloc(mem_size_C);
 
     checkCudaErrors(cudaMalloc((void **) &d_A, mem_size_A));
     checkCudaErrors(cudaMalloc((void **) &d_B, mem_size_B));
@@ -243,19 +292,30 @@ int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
     printf("Computing result using CUBLAS...");
 
     // execute the kernel
-    int nIter = 30;
+    int nIter = 9999999;
 
     // CUBLAS version 2.0
     {
-        const float alpha = 1.0f;
-        const float beta  = 0.0f;
+        cublasOperation_t trans_A = (ta) ? CUBLAS_OP_T : CUBLAS_OP_N;
+        cublasOperation_t trans_B = (tb) ? CUBLAS_OP_T : CUBLAS_OP_N;
+        int m = matrix_size.uiWC;
+        int n = matrix_size.uiHC;
+        int k = matrix_size.uiWA;
+        int lda = (trans_A == CUBLAS_OP_N) ? k : n;
+        int ldb = (trans_B == CUBLAS_OP_N) ? m : k;
+        int ldc = m;
+        const __half alpha = 1.0f;
+        const __half beta  = 0.0f;
         cublasHandle_t handle;
         cudaEvent_t start, stop;
 
         checkCudaErrors(cublasCreate(&handle));
 
         //Perform warmup operation with cublas
-        checkCudaErrors(cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, matrix_size.uiWB, matrix_size.uiHA, matrix_size.uiWA, &alpha, d_B, matrix_size.uiWB, d_A, matrix_size.uiWA, &beta, d_C, matrix_size.uiWB));
+        checkCudaErrors(cublasSetMathMode(handle, CUBLAS_TENSOR_OP_MATH));
+        cublasGemmAlgo_t algo = cublasGemmAlgo_t::CUBLAS_GEMM_ALGO4_TENSOR_OP;
+        checkCudaErrors(cublasGemmEx(handle, trans_B, trans_A, m, n, k, &alpha, d_B, CUDA_R_16F, 
+                    ldb, d_A, CUDA_R_16F, lda, &beta, d_C, CUDA_R_16F, ldc, CUBLAS_COMPUTE_16F, algo));
 
         // Allocate CUDA events that we'll use for timing
         checkCudaErrors(cudaEventCreate(&start));
@@ -268,8 +328,8 @@ int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
         {
             //note cublas is column primary!
             //need to transpose the order
-            checkCudaErrors(cublasSgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, matrix_size.uiWB, matrix_size.uiHA, matrix_size.uiWA, &alpha, d_B, matrix_size.uiWB, d_A, matrix_size.uiWA, &beta, d_C, matrix_size.uiWB));
-
+            checkCudaErrors(cublasGemmEx(handle, trans_B, trans_A, m, n, k, &alpha, d_B, CUDA_R_16F, 
+                        ldb, d_A, CUDA_R_16F, lda, &beta, d_C, CUDA_R_16F, ldc, CUBLAS_COMPUTE_16F, algo));
         }
 
         printf("done.\n");
@@ -302,16 +362,16 @@ int matrixMultiply(int argc, char **argv, int devID, sMatrixSize &matrix_size)
 
     // compute reference solution
     printf("Computing result using host CPU...");
-    float *reference = (float *)malloc(mem_size_C);
-    matrixMulCPU(reference, h_A, h_B, matrix_size.uiHA, matrix_size.uiWA, matrix_size.uiWB);
+    __half *reference = (__half *)malloc(mem_size_C);
+    matrixMulCPU(reference, h_A, h_B, matrix_size.uiHA, matrix_size.uiWA, matrix_size.uiWB, ta, tb);
     printf("done.\n");
 
     // check result (CUBLAS)
-    bool resCUBLAS = sdkCompareL2fe(reference, h_CUBLAS, size_C, 1.0e-6f);
+    bool resCUBLAS = sdkCompareL2fe(reference, h_CUBLAS, size_C, 1.0e-2f);
 
     if (resCUBLAS != true)
     {
-        printDiff(reference, h_CUBLAS, matrix_size.uiWC, matrix_size.uiHC, 100, 1.0e-5f);
+        printDiff(reference, h_CUBLAS, matrix_size.uiWC, matrix_size.uiHC, 100, 1.0e-2f);
     }
 
     printf("Comparing CUBLAS Matrix Multiply with CPU results: %s\n", (true == resCUBLAS) ? "PASS" : "FAIL");
@@ -349,7 +409,7 @@ int main(int argc, char **argv)
 
     initializeCUDA(argc, argv, devID, sizeMult, matrix_size);
 
-    int matrix_result = matrixMultiply(argc, argv, devID, matrix_size);
+    int matrix_result = matrixMultiply(argc, argv, devID, matrix_size, true, false);
 
     return matrix_result;
 }

3. Run the sample

$ ./matrixMulCUBLAS


Installation Steps

Darknet with cuDNN-8 Support

Below are the steps to build darknet with cuDNN-8 support.

Verified environment:

  • JetPack4.5.1 + Xavier

1. Get source

$ git clone https://github.com/pjreddie/darknet.git
$ cd darknet/
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/script/topics/0001-fix-for-cudnn_v8-limited-memory-to-default-darknet-s.patch
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/opencv-darknet.patch -O opencv-darknet.patch
$ git am 0001-fix-for-cudnn_v8-limited-memory-to-default-darknet-s.patch
$ git am opencv-darknet.patch

2. Update Makefile based on your device

GPU=1
CUDNN=1
OPENCV=1
  • Xavier & XavierNX:
ARCH= -gencode arch=compute_72,code=sm_72 \
      -gencode arch=compute_72,code=[sm_72,compute_72]
  • TX2:
ARCH= -gencode arch=compute_62,code=sm_62 \
      -gencode arch=compute_62,code=[sm_62,compute_62]
  • Nano:
ARCH= -gencode arch=compute_53,code=sm_53 \
      -gencode arch=compute_53,code=[sm_53,compute_53]

3. Build and Test

$ make -j8
$ wget https://pjreddie.com/media/files/yolov3-tiny.weights
$ ./darknet detector demo cfg/coco.data cfg/yolov3-tiny.cfg yolov3-tiny.weights [video]


TensorRT Python Bindings

Below are the steps to build TensorRT Python 3.9 bindings.

Verified environment:

  • JetPack4.6 + Xavier

1. Building python3.9

$ sudo apt install zlib1g-dev libncurses5-dev libgdbm-dev libnss3-dev libssl-dev libreadline-dev libffi-dev libsqlite3-dev libbz2-dev
$ wget https://www.python.org/ftp/python/3.9.1/Python-3.9.1.tar.xz
$ tar xvf Python-3.9.1.tar.xz Python-3.9.1/
$ mkdir build-python-3.9.1
$ cd build-python-3.9.1/
$ ../Python-3.9.1/configure --enable-optimizations
$ make -j $(nproc)
$ sudo -H make altinstall
$ cd ../

2. Build cmake 3.13.5

$ sudo apt-get install -y protobuf-compiler libprotobuf-dev openssl libssl-dev libcurl4-openssl-dev
$ wget https://github.com/Kitware/CMake/releases/download/v3.13.5/cmake-3.13.5.tar.gz
$ tar xvf cmake-3.13.5.tar.gz
$ rm cmake-3.13.5.tar.gz
$ cd cmake-3.13.5/
$ ./bootstrap --system-curl
$ make -j$(nproc)
$ echo 'export PATH='${PWD}'/bin/:$PATH' >> ~/.bashrc
$ source ~/.bashrc
$ cd ../

3. Prepare header

$ mkdir python3.9
$ mkdir python3.9/include
$ wget http://ftp.us.debian.org/debian/pool/main/p/python3.9/libpython3.9-dev_3.9.9-2_arm64.deb
$ ar x libpython3.9-dev_3.9.9-2_arm64.deb
$ tar -xvf data.tar.xz
$ cp ./usr/include/aarch64-linux-gnu/python3.9/pyconfig.h python3.9/include/
$ cp -r Python-3.9.1/Include/* python3.9/include/

4. Build TensorRT pybinding

$ git clone https://github.com/pybind/pybind11.git
$ git clone -b release/8.0 https://github.com/NVIDIA/TensorRT.git
$ cd TensorRT
$ git submodule update --init --recursive
$ cd python/
$ TRT_OSSPATH=${PWD}/.. EXT_PATH=${PWD}/../.. TARGET=aarch64 PYTHON_MINOR_VERSION=9 ./build.sh
$ python3.9 -m pip install build/dist/tensorrt-8.0.1.6-cp39-none-linux_aarch64.whl


Caffe

Below are the steps to build the Caffe library.

Verified environment:

  • JetPack4.6 + Xavier
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/install_caffe_jp46.sh -O install_caffe_jp46.sh
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/0001-patch-for-jp4.6.patch -O 0001-patch-for-jp4.6.patch
$ ./install_caffe_jp46.sh
$ source ~/.bashrc


MXNet

Below are the steps to build the MXNet 1.8.0 library.

Verified environment:

  • JetPack4.5.1 + Xavier
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/mxnet_v1.8.x.patch -O mxnet_v1.8.x.patch
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/autobuild_mxnet.sh -O autobuild_mxnet.sh
$ sudo chmod +x autobuild_mxnet.sh
$ ./autobuild_mxnet.sh Xavier
$ cd mxnet/build/
$ pip3 install mxnet-1.8.0-py3-none-any.whl


PyInstaller with OpenCV

Currently, the OpenCV version between JetPack default and Pyinstaller is not consistent.

To solve this issue, you can either upgrade the python-opencv version or downgrade the PyInstaller version.

  • Upgrade python-opencv
$ pip3 install opencv-python
  • Downgrade pyinstaller and pyinstaller-hooks-contrib
$ sudo pip3 install pyinstaller==4.2
$ sudo pip3 install pyinstaller-hooks-contrib==2021.2
$ pyinstaller --onefile --paths="/usr/lib/python3.6/dist-packages/cv2/python-3.6" myfile.py


Common Issues

"Unsupported ONNX data type: UINT8 (2)"

This error is from TensorRT. The root cause is that ONNX expects the input image to be INT8 but TensorRT uses Float32.

To solve this issue, you can modify the input data format of ONNX with our graphsurgeon API.

$ sudo apt-get install python3-pip libprotobuf-dev protobuf-compiler
$ git clone https://github.com/NVIDIA/TensorRT.git
$ cd TensorRT/tools/onnx-graphsurgeon/
$ make install
import onnx_graphsurgeon as gs
import onnx
import numpy as np

graph = gs.import_onnx(onnx.load("model.onnx"))
for inp in graph.inputs:
    inp.dtype = np.float32

onnx.save(gs.export_onnx(graph), "updated_model.onnx")


"Illegal instruction (core dumped)"

This is a known issue in NumPy v1.19.5.

To solve this issue, you can either downgrade your NumPy into 1.19.4 or manually update an environment variable.

  • Downgrade NumPy
$ sudo apt-get install python3-pip
$ pip3 install Cython
$ pip3 install numpy==1.19.4
  • Update environment variable
$ export OPENBLAS_CORETYPE=ARMV8


Long delays when submitting several cudaMemcpy

Please try to increase the computing channel

$ export CUDA_DEVICE_MAX_CONNECTIONS=32

A document can be found here:

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars