| 203 | } |
| 204 | |
| 205 | template <typename ElemType> int |
| 206 | HemmPerformanceTest<ElemType>::prepare(void) |
| 207 | { |
| 208 | //bool useAlpha = base_->useAlpha(); |
| 209 | //bool useBeta = base_->useBeta(); |
| 210 | |
| 211 | int creationFlags = 0, AcreationFlags = 0; |
| 212 | creationFlags = creationFlags | RANDOM_INIT; |
| 213 | |
| 214 | creationFlags = ( (params_.order) == clblasRowMajor)? (creationFlags | ROW_MAJOR_ORDER) : (creationFlags); |
| 215 | AcreationFlags = ( (params_.uplo) == clblasLower)? (creationFlags | LOWER_HALF_ONLY) : (creationFlags | UPPER_HALF_ONLY); |
| 216 | BlasRoutineID BlasFn = CLBLAS_HEMM; |
| 217 | |
| 218 | populate( A_ + params_.offA, ka, ka, params_.lda, BlasFn, AcreationFlags); |
| 219 | populate( B_ + params_.offBX, params_.M, params_.N, params_.ldb, BlasFn, creationFlags ); |
| 220 | populate( C_ + params_.offCY, params_.M, params_.N, params_.ldc, BlasFn, creationFlags ); |
| 221 | |
| 222 | memcpy( backC_, C_, (kbc * params_.ldc + params_.offCY) * sizeof(ElemType) ); |
| 223 | |
| 224 | mobjA_ = base_->createEnqueueBuffer(A_, (params_.lda * ka + params_.offA) * sizeof(ElemType), 0, CL_MEM_READ_ONLY); |
| 225 | mobjB_ = base_->createEnqueueBuffer(B_, (params_.ldb * kbc + params_.offBX) * sizeof(ElemType), 0, CL_MEM_READ_ONLY); |
| 226 | mobjC_ = base_->createEnqueueBuffer(backC_, (params_.ldc * kbc + params_.offCY) * sizeof(ElemType), 0, CL_MEM_READ_WRITE); |
| 227 | |
| 228 | return (mobjC_) ? 0 : -1; |
| 229 | } |
| 230 | |
| 231 | template <typename ElemType> nano_time_t |
| 232 | HemmPerformanceTest<ElemType>::etalonPerfSingle(void) |
nothing calls this directly
no test coverage detected