Optimizing OpenCL Kernels for the ARM Mali-T600 GPUs - GPU Pro: Advanced Rendering Techniques - page 335

kernel void sobel_char8 (
global const uchar ￿ restrict in , // < Input .
global char ￿ restrict dx ,

// < Xgradient.

global char ￿ restrict dy , // < Ygradient.

const int width )
{ // X and Y gradient accumulators .
short8 _dx , _dy ;

// Left , middle and right pixels : loaded as
// unsigned 8

− bit , converted to signed 16 − bit .

uchar8 lLoad , mLoad , rLoad ;
short8 lData , mData , rData ;

// Compute (column , row) position and off set .
const int column = get global id (0) ￿ 8;
const int row = get global id (1) ￿ 1;
const int offset = row ￿ width + column ;
// Compute contribution from f i r s t row .
lLoad = vload8 (0, in +( offset + width ￿ 0+0));
mLoad = vload8 (0, in +( offset + width ￿ 0+1));
rLoad = vload8 (0, in +( offset + width ￿ 0+2));
lData = convert short8 ( lLoad );
mData = convert short8 ( mLoad );
rData = convert short8 ( rLoad );

_dx = rData − lData ;
_dy = rData + lData + mData ￿ ( short8 )2;
// Compute contribution from second row .
lLoad = vload8 (0, in +( offset + width ￿ 1+0));
rLoad = vload8 (0, in +( offset + width ￿ 1+2));
lData = convert short8 ( lLoad );
rData = convert short8 ( rLoad );

_dx += ( rData − lData ) ￿ ( short8 )2;
// Compute contribution from third row.
lLoad = vload8 (0, in +( offset + width ￿ 2+0));
mLoad = vload8 (0, in +( offset + width ￿ 2+1));
rLoad = vload8 (0, in +( offset + width ￿ 2+2));
lData = convert short8 ( lLoad );
mData = convert short8 ( mLoad );
rData = convert short8 ( rLoad );

_dx += rData − lData ;
_dy − = rData + lData + mData

( short8 )2;

// Store the results .
vstore8 ( convert char8 ( _dx >> 3) , 0, dx + offset + width +1);
vstore8 ( convert char8 ( _dy >> 3) , 0, dy + offset + width +1);

}

Listing 7.2. Initial vector implementation: char8 .