| 28 | namespace kernel { |
| 29 | template<typename T> |
| 30 | void cscmv(Param out, const Param &values, const Param &colIdx, |
| 31 | const Param &rowIdx, const Param &rhs, const T alpha, const T beta, |
| 32 | bool is_conj) { |
| 33 | // TODO: rows_per_group limited by register pressure. Find better way to |
| 34 | // handle this. |
| 35 | constexpr int threads_per_g = 64; |
| 36 | constexpr int rows_per_group = 64; |
| 37 | |
| 38 | const bool use_alpha = (alpha != scalar<T>(1.0)); |
| 39 | const bool use_beta = (beta != scalar<T>(0.0)); |
| 40 | |
| 41 | cl::NDRange local(threads_per_g); |
| 42 | |
| 43 | int K = colIdx.info.dims[0] - 1; |
| 44 | int M = out.info.dims[0]; |
| 45 | |
| 46 | std::array<TemplateArg, 5> targs = { |
| 47 | TemplateTypename<T>(), TemplateArg(use_alpha), |
| 48 | TemplateArg(is_conj), TemplateArg(rows_per_group), |
| 49 | TemplateArg(local[0]), |
| 50 | }; |
| 51 | std::array<std::string, 9> options = { |
| 52 | DefineKeyValue(T, dtype_traits<T>::getName()), |
| 53 | DefineKeyValue(USE_ALPHA, use_alpha), |
| 54 | DefineKeyValue(IS_CONJ, is_conj), |
| 55 | DefineKeyValue(THREADS, local[0]), |
| 56 | DefineKeyValue(ROWS_PER_GROUP, rows_per_group), |
| 57 | DefineKeyValue(IS_CPLX, (iscplx<T>() ? 1 : 0)), |
| 58 | DefineKeyValue(IS_DBL, (isdbl<T>() ? 1 : 0)), |
| 59 | DefineKeyValue(IS_LONG, (islong<T>() ? 1 : 0)), |
| 60 | getTypeBuildDefinition<T>()}; |
| 61 | |
| 62 | if(use_beta) { |
| 63 | std::array<TemplateArg, 4> targs_beta = { |
| 64 | TemplateTypename<T>(), TemplateArg(is_conj), |
| 65 | TemplateArg(rows_per_group), TemplateArg(local[0])}; |
| 66 | std::array<std::string, 8> options_beta = { |
| 67 | DefineKeyValue(T, dtype_traits<T>::getName()), |
| 68 | DefineKeyValue(IS_CONJ, is_conj), |
| 69 | DefineKeyValue(THREADS, local[0]), |
| 70 | DefineKeyValue(ROWS_PER_GROUP, rows_per_group), |
| 71 | DefineKeyValue(IS_CPLX, (iscplx<T>() ? 1 : 0)), |
| 72 | DefineKeyValue(IS_DBL, (isdbl<T>() ? 1 : 0)), |
| 73 | DefineKeyValue(IS_LONG, (islong<T>() ? 1 : 0)), |
| 74 | getTypeBuildDefinition<T>()}; |
| 75 | |
| 76 | int groups_x = divup(M, rows_per_group * threads_per_g); |
| 77 | cl::NDRange global(local[0] * groups_x, 1); |
| 78 | auto cscmvBeta = common::getKernel("cscmv_beta", {{cscmv_cl_src}}, targs_beta, options_beta); |
| 79 | cscmvBeta(cl::EnqueueArgs(getQueue(), global, local), *out.data, M, beta); |
| 80 | |
| 81 | } else { |
| 82 | getQueue().enqueueFillBuffer(*out.data, 0, 0, M * sizeof(T)); |
| 83 | } |
| 84 | |
| 85 | int groups_x = divup(M, rows_per_group); |
| 86 | cl::NDRange global(local[0] * groups_x, 1); |
| 87 |
no test coverage detected