OpenCL does not take unicode strings as input, so this routine returns only ASCII strings Feed this generator the FFTPlan, and it returns the generated program as a string
| 197 | // OpenCL does not take unicode strings as input, so this routine returns only ASCII strings |
| 198 | // Feed this generator the FFTPlan, and it returns the generated program as a string |
| 199 | clfftStatus FFTGeneratedTransposeNonSquareAction::generateKernel(FFTRepo& fftRepo, const cl_command_queue commQueueFFT) |
| 200 | { |
| 201 | |
| 202 | |
| 203 | std::string programCode; |
| 204 | std::string kernelFuncName;//applied to swap kernel for now |
| 205 | if (this->signature.nonSquareKernelType == NON_SQUARE_TRANS_TRANSPOSE_BATCHED_LEADING) |
| 206 | { |
| 207 | //Requested local memory size by callback must not exceed the device LDS limits after factoring the LDS size required by transpose kernel |
| 208 | if (this->signature.fft_hasPreCallback && this->signature.fft_preCallback.localMemSize > 0) |
| 209 | { |
| 210 | assert(!this->signature.fft_hasPostCallback); |
| 211 | |
| 212 | bool validLDSSize = false; |
| 213 | size_t requestedCallbackLDS = 0; |
| 214 | |
| 215 | requestedCallbackLDS = this->signature.fft_preCallback.localMemSize; |
| 216 | |
| 217 | validLDSSize = ((2 * this->plan->ElementSize() * 16 * reShapeFactor * 16 * reShapeFactor) + requestedCallbackLDS) < this->plan->envelope.limit_LocalMemSize; |
| 218 | |
| 219 | if(!validLDSSize) |
| 220 | { |
| 221 | fprintf(stderr, "Requested local memory size not available\n"); |
| 222 | return CLFFT_INVALID_ARG_VALUE; |
| 223 | } |
| 224 | } |
| 225 | OPENCL_V(clfft_transpose_generator::genTransposeKernelLeadingDimensionBatched(this->signature, programCode, lwSize, reShapeFactor), _T("genTransposeKernel() failed!")); |
| 226 | } |
| 227 | else if (this->signature.nonSquareKernelType == NON_SQUARE_TRANS_TRANSPOSE_BATCHED) |
| 228 | { |
| 229 | //pre call back check |
| 230 | //Requested local memory size by callback must not exceed the device LDS limits after factoring the LDS size required by transpose kernel |
| 231 | if (this->signature.fft_hasPreCallback && this->signature.fft_preCallback.localMemSize > 0) |
| 232 | { |
| 233 | assert(!this->signature.fft_hasPostCallback); |
| 234 | |
| 235 | bool validLDSSize = false; |
| 236 | size_t requestedCallbackLDS = 0; |
| 237 | |
| 238 | requestedCallbackLDS = this->signature.fft_preCallback.localMemSize; |
| 239 | |
| 240 | validLDSSize = ((2 * this->plan->ElementSize() * 16 * reShapeFactor * 16 * reShapeFactor) + requestedCallbackLDS) < this->plan->envelope.limit_LocalMemSize; |
| 241 | |
| 242 | if (!validLDSSize) |
| 243 | { |
| 244 | fprintf(stderr, "Requested local memory size not available\n"); |
| 245 | return CLFFT_INVALID_ARG_VALUE; |
| 246 | } |
| 247 | } |
| 248 | OPENCL_V(clfft_transpose_generator::genTransposeKernelBatched(this->signature, programCode, lwSize, reShapeFactor), _T("genTransposeKernel() failed!")); |
| 249 | } |
| 250 | else |
| 251 | { |
| 252 | //pre-callback is possible in swap kernel now |
| 253 | if (this->signature.fft_hasPreCallback && this->signature.fft_preCallback.localMemSize > 0) |
| 254 | { |
| 255 | assert(!this->signature.fft_hasPostCallback); |
| 256 |
no test coverage detected