cudaMemcpyToSymbol reversing order representation of byte array

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I’m new to CUDA and I’m losing my mind over a strange behaviour of cudaMemcpyToSymbol.

I need to copy an array of 20 unsigned char (bHash_h) from the host into 5 unsigned long (bHash) in the device constant memory.

The code is:

[codebox]unsigned char bHash_h[20] = { 0x62, 0x6D, 0x36, 0xE7, 0x11, 0x78, 0x26, 0xCF, 0xF4, 0xD6, 0xC5, 0xB5, 0x66, 0x9B, 0x08, 0xAF, 0xD9, 0x95, 0x14, 0x9A};

int main(int argc, char* argv)

{

…

rc = cudaMemcpyToSymbol( "bHash", bHash_h, 20);

}

/******* device code ***********/

constant unsigned long bHash[5];

global void myFunc( unsigned long *Result )

{

…

printf(" bhash[0] %08X \n", bHash[0]);

…

}

[/codebox]

The output is :

E7366D62

instead of:

626D36E7

How can it be?

I tried to use

rc = cudaMemcpyToSymbol( “bHash”, bHash_h, 20, 0, cudaMemcpyHostToDevice);

but I didn’t work…

Thanks for your time in advance!!

Martina

2

The output is correct. The IA-32/x86_64 are little endian architectures - the first byte in a word is the least significant byte and the last is the most signficant byte. You can read more here if you haven’t come across the concept before.

3

ups it is an endianess issue… Thank you!! so I guess I need to swap the bytes or is there another (faster) method?