Which is the best way to check a bit array in Cuda? - cuda

Question

I need to launch N threads (in one block)
This is the code, 'e' is a bignumber on 1024b. I need to copy it on the gpu and read it bit by bit.
Host code:
unsigned char *__e;
BIGNUM *e = BN_new();
unsigned char exp[128];
// e
i = cudaMalloc( (void**)&__e, 128* sizeof(unsigned char) );
if(i != cudaSuccess)
printf("cudaMalloc __e FAIL! Code: %d\n", i);
BN_bn2bin128B(e, exp); // copy data in exp
for(i=0; i<128; i++)
exp[i] = reverse(exp[i]);
i = cudaMemcpy( __e, exp, 128* sizeof(unsigned char), cudaMemcpyHostToDevice);
if(i != cudaSuccess)
printf("cudaMemcpy __e FAIL! Code: %d\n", i);
unsigned char reverse(unsigned char b) {
b = (b & 0xF0) >> 4 | (b & 0x0F) << 4;
b = (b & 0xCC) >> 2 | (b & 0x33) << 2;
b = (b & 0xAA) >> 1 | (b & 0x55) << 1;
return b;
}
Device code:
for(int i=0; i<1024; i++)
if(ISBITSET(__e, i) == 1)
//do something
Header:
#define ISBITSET(x,i) ((x[i>>3] & (1<<(i&7)))!=0)
Unfortunately ISBITSET doesnt accept anything different from __e, so I cant check the further values in __e itself
How can I solve it? Or is there a better way?

Answer 1

The GPU is a 32 bit machine, so you'll want to process your 1024 bits 32 bits at a time, not 8. So, you should replace all unsigned char with unsigned int and adjust the values accordingly.
The GPU has a fast PTX instruction for reversing 32 bits at a time, so you may want to implement that on the GPU. The instruction is called brev. To use it, you would add inline PTX, something like this (untested):
asm("brev.b32 %0, %1;" : "=r"(dst_var) : "r"(src_var));
For more information, see NVIDIA's document, "Using Inline PTX Assembly In CUDA".
for(int i=0; i<1024; i++)
if(ISBITSET(__e, i) == 1)
//do something
This code may have a performance problem. Presuming that there is a 50% chance that a bit is on, you only get 50% of the possible performance since half of your threads will have to wait while the other half perform the //do something. I can't think of a workaround though. You may also want to launch threads instead of looping.
Unfortunately ISBITSET doesnt accept anything different from __e, so I cant check the further values in __e itself
Could you elaborate? The ISBITSET macro looks ok to me and looks like it can process any array of unsigned chars, which is what __e is.