time execution HIGHT and LOW !?!?!?

Accelerated Computing CUDA CUDA Programming and Performance
1

My very problem is this:

example:

global void kernel(unsigned short* dataDevice,uchar4* output)

{

int j = (blockIdx.x) * blockDim.x + threadIdx.x;		

int i = (blockIdx.y) * blockDim.y + threadIdx.y;	

if ((j<width) && (i<height))

{

float luminance;

            float variable;

            ........

             ......

              calculation a corretct index

            ....

            ....

            for (int k = 0; k<index; k++)

            {

                     luminance+=(float)dataDevice[........];

variable += 1; (for example)

            }

…

           ......

…

output[i*width+j] = make_uchar4(luminance,luminance,luminance,255);

    }

}

time execution VERY VERY VERY HIGHT

identical except “output[i*width+j] = make_uchar4(luminance,luminance,luminance,255);” but for example output[i*width+j] = make_uchar4(100,100,100,255);" or another variable but not luminance

“output[i*width+j] = make_uchar4(lvariable,variable,variable,255);” (for example)

time execution VERY VERY VERY LOW

why? read/write whit luminance? wait ? conflict? WHY?

2

My very problem is this:

example:

global void kernel(unsigned short* dataDevice,uchar4* output)

{

int j = (blockIdx.x) * blockDim.x + threadIdx.x;		

int i = (blockIdx.y) * blockDim.y + threadIdx.y;	

if ((j<width) && (i<height))

{

float luminance;

            float variable;

            ........

             ......

              calculation a corretct index

            ....

            ....

            for (int k = 0; k<index; k++)

            {

                     luminance+=(float)dataDevice[........];

variable += 1; (for example)

            }

…

           ......

…

output[i*width+j] = make_uchar4(luminance,luminance,luminance,255);

    }

}

time execution VERY VERY VERY HIGHT

identical except “output[i*width+j] = make_uchar4(luminance,luminance,luminance,255);” but for example output[i*width+j] = make_uchar4(100,100,100,255);" or another variable but not luminance

“output[i*width+j] = make_uchar4(lvariable,variable,variable,255);” (for example)

time execution VERY VERY VERY LOW

why? read/write whit luminance? wait ? conflict? WHY?

3

The compiler is able to optimize away the whole luminance calculation if it’s result is not used.

4

The compiler is able to optimize away the whole luminance calculation if it’s result is not used.

5

for example: nThreads (32,16) nBlock(200,400) with luminance ~ 8000 ms !!! (I disabled the “Timeout Detection and Recovery”)

without luminance ~ 0.002 ms

6

for example: nThreads (32,16) nBlock(200,400) with luminance ~ 8000 ms !!! (I disabled the “Timeout Detection and Recovery”)

without luminance ~ 0.002 ms