---

Error handling in CUDA kernel launch

---

• Error handling
  • Return behavior of every CUDA function:

    If a CUDA function completes normally (= no error) then: The function will return the value cudaSuccess

    ---

  • cudaGetLastError( ):

    The CDUA library function cudaGetLastError( ) returns: The error code of the last executed CUDA function

    ---

  • Commonly used error handling code in CUDA programming:

    call a CUDA kernel (or launch a kernel); cudaError_t err = cudaGetLastError(); if ( err != cudaSuccess ) { printf("CUDA Error: %s\n", cudaGetErrorString(err)); // Possibly: exit(-1) if program cannot continue.... }

  ---

  ---

  ---

  ---

• Advice: always check the return value of CUDA functions !!!
  • Example CUDA program with execution error:

    __global__ void hello( ) { printf("blockIdx.x=%d/%d blocks, threadIdx.x=%d/%d threads\n", blockIdx.x, gridDim.x, threadIdx.x, blockDim.x); } int main() { hello<<< 1, 1025 >>>( ); // Error: #threads >= 1024 !!! printf("I am the CPU: Hello World ! \n"); cudaDeviceSynchronize(); return 0; }

    Warning:

    CUDA execution error are often not visible when you run the program !!!

  • Example Program: (Demo above code)                                                

    /home/cs355001/demo/CUDA/1-intro/hello-error Output: I am the CPU: Hello World ! (No error message from the GPU execution !!!)

    ---

    ---

  • Better code:

    __global__ void hello( ) { printf("blockIdx.x=%d/%d blocks, threadIdx.x=%d/%d threads\n", blockIdx.x, gridDim.x, threadIdx.x, blockDim.x); } int main() { hello<<< 1, 1025 >>>( ); // Error !!! cudaError_t err = cudaGetLastError(); // Get error code if ( err != cudaSuccess ) { printf("CUDA Error: %s\n", cudaGetErrorString(err)); exit(-1); } printf("I am the CPU: Hello World ! \n"); cudaDeviceSynchronize(); return 0; }

  • Example Program: (Demo above code)                                                

    /home/cs355001/demo/CUDA/1-intro/hello-error2 Output: CUDA Error: invalid configuration argument