FIXME: Report correct return value - Needs change in KPRINTF
| 209 | // FIXME: Report correct return value - Needs change in KPRINTF |
| 210 | // |
| 211 | static ssize_t |
| 212 | generator( |
| 213 | char *buf, |
| 214 | size_t buflen, |
| 215 | const struct SubproblemDim *subdims, |
| 216 | const struct PGranularity *pgran, |
| 217 | void *extra) |
| 218 | { |
| 219 | |
| 220 | DUMMY_ARG_USAGE(subdims); |
| 221 | size_t BLOCKSIZE = pgran->wgSize[0]; |
| 222 | char tempTemplate[32*1024]; |
| 223 | |
| 224 | if ( buf == NULL) // return buffer size |
| 225 | { |
| 226 | buflen = (32 * 1024 * sizeof(char)); |
| 227 | return (ssize_t)buflen; |
| 228 | } |
| 229 | CLBLASKernExtra *extraFlags = ( CLBLASKernExtra *)extra; |
| 230 | |
| 231 | #ifdef DEBUG_ASUM |
| 232 | printf("ASUM GENERATOR called....\n"); |
| 233 | printf("dataType : %c\n", Prefix[extraFlags->dtype]); |
| 234 | #endif |
| 235 | |
| 236 | unsigned int vecLenA = extraFlags->vecLenA; |
| 237 | |
| 238 | #ifdef DEBUG_ASUM |
| 239 | printf("Vector length used : %d\n\n", vecLenA); |
| 240 | #endif |
| 241 | |
| 242 | bool doVLOAD = false; |
| 243 | if( extraFlags->flags & KEXTRA_NO_COPY_VEC_A ) |
| 244 | { |
| 245 | doVLOAD = true; |
| 246 | #ifdef DEBUG_ASUM |
| 247 | printf("DOing VLOAD as Aligned Data Pointer not Availabe\n"); |
| 248 | #endif |
| 249 | } |
| 250 | else |
| 251 | { |
| 252 | #ifdef DEBUG_ASUM |
| 253 | printf("Using Aligned Data Pointer \n"); |
| 254 | #endif |
| 255 | } |
| 256 | strcpy( tempTemplate, (char*)asum_kernel ); |
| 257 | kprintf kobj( Prefix[extraFlags->dtype], vecLenA, doVLOAD, doVLOAD, BLOCKSIZE); |
| 258 | kobj.spit((char*)buf, tempTemplate); |
| 259 | |
| 260 | return (32 * 1024 * sizeof(char)); |
| 261 | } |
| 262 | |
| 263 | /* |
| 264 | __kernel void %PREFIXasum_kernel( __global %TYPE *_X, __global %TYPE *scratchBuff, uint N, uint offx, int incx) |