CUDA allocating array of strings

Question

Whenever I run my script that contains those lines:

   char ** gpu_reads;
   HANDLE_ERROR(cudaMalloc((void **)&gpu_reads, inputDim * sizeof(char *)));
   for(i=0; i<inputDim; i++) {
      HANDLE_ERROR(cudaMalloc((void **)&(gpu_reads[i]), (READS_LENGTH + 1) * sizeof(char)));
   }
   for(i=0; i<inputDim; i++) {
      HANDLE_ERROR(cudaMemcpy(gpu_reads[i], reads[i], sizeof(char) * (READS_LENGTH + 1), cudaMemcpyHostToDevice));
   }

The second line returns an "unknown error". I tried different allocation in my program (this was the first one) but none of them worked.

The purpose of those lines is simply allocating an array (of length fixed by user, using the variable inputDim of strings (of fixed length).

I tried different version (i.e. using only 3 pointers, 1 pointer...) but none seems to work...

Any ideas?

The full code is available at my GitHub repository, where many allocation are made.

Why do have the same number of stars on the 4th line as the 2nd in (void **)? — Weather Vane
– Weather Vane, Commented Jun 9, 2016 at 18:41
Allocate (READS_LENGTH * inputDim) bytes in a single chunk and you will never have to struggle with broken loops. — Drop
– Drop, Commented Jun 10, 2016 at 6:28

talonmies · Accepted Answer · 2016-06-09 18:54:30Z

3

What you are trying to do cannot work because your code attempts to access memory you have allocated on the device from the host. You cannot access the elements of gpu_readson the host because it is not a valid host memory allocation.

Do something like this instead:

   char ** gpu_reads;
   char ** gpu_reads_h = new char*[input_dim];
   HANDLE_ERROR(cudaMalloc((void **)&gpu_reads, inputDim * sizeof(char *)));
   for(i=0; i<inputDim; i++) {
      HANDLE_ERROR(cudaMalloc((void **)&(gpu_reads_h[i]), (READS_LENGTH + 1) * sizeof(char)));
   }
   for(i=0; i<inputDim; i++) {
      HANDLE_ERROR(cudaMemcpy(gpu_reads_h[i], reads[i], sizeof(char) * (READS_LENGTH + 1), cudaMemcpyHostToDevice));
   }

   HANDLE_ERROR(cudaMemcpy(gpu_reads, gpu_reads_h, inputDim * sizeof(char *), cudaMemcpyHostToDevice);

i.e. build a copy of the eventual device array of pointers in host memory first, then copy it to the device.

answered Jun 9, 2016 at 18:54

talonmies

72.8k35 gold badges204 silver badges296 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

user6446972 Over a year ago

I tried this code, but I have the same "unknown error"... I also tried to allocate with (READS_LENGTH * inputDim) like said @Drop but it doesn't work either

talonmies Over a year ago

@Cordaz: If you are getting that error then you have another problem unrelated to the memory allocation and copying, like a non-functioning CUDA installation, or prior broken code which you haven't shown us. Here is a complete example from your code snippet which works perfectly for me. Try it for yourself

user6446972 Over a year ago

I tried your code and I get the same error... I installed CUDA 7.5 o Ubuntu 16.04. I tried running it with sudo, and it returned this line: modprobe: FATAL: Module nvidia-uvm not found in directory /lib/modules/4.4.0-22-generic I already tried to install this package, but after the reboot the system started in low graphic mode and I had to restore previous Nvidia driver...

talonmies Over a year ago

OK so you have a broken CUDA installation. Stack Overflow isn't the correct place to get help with that, I am afraid.

Collectives™ on Stack Overflow

CUDA allocating array of strings

1 Answer 1

4 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Related