| 29 | namespace kernel { |
| 30 | template<typename T> |
| 31 | void csrmv(Param out, const Param &values, const Param &rowIdx, |
| 32 | const Param &colIdx, const Param &rhs, const T alpha, const T beta) { |
| 33 | constexpr int MAX_CSRMV_GROUPS = 4096; |
| 34 | // Using greedy indexing is causing performance issues on many platforms |
| 35 | // FIXME: Figure out why |
| 36 | constexpr bool use_greedy = false; |
| 37 | |
| 38 | // TODO: Figure out the proper way to choose either csrmv_thread or |
| 39 | // csrmv_block |
| 40 | bool is_csrmv_block = true; |
| 41 | |
| 42 | const bool use_alpha = (alpha != scalar<T>(1.0)); |
| 43 | const bool use_beta = (beta != scalar<T>(0.0)); |
| 44 | |
| 45 | cl::NDRange local(THREADS_PER_GROUP); |
| 46 | |
| 47 | std::array<TemplateArg, 5> targs = { |
| 48 | TemplateTypename<T>(), TemplateArg(use_alpha), TemplateArg(use_beta), |
| 49 | TemplateArg(use_greedy), TemplateArg(local[0]), |
| 50 | }; |
| 51 | std::array<std::string, 7> options = { |
| 52 | DefineKeyValue(T, dtype_traits<T>::getName()), |
| 53 | DefineKeyValue(USE_ALPHA, use_alpha), |
| 54 | DefineKeyValue(USE_BETA, use_beta), |
| 55 | DefineKeyValue(USE_GREEDY, use_greedy), |
| 56 | DefineKeyValue(THREADS, local[0]), |
| 57 | DefineKeyValue(IS_CPLX, (iscplx<T>() ? 1 : 0)), |
| 58 | getTypeBuildDefinition<T>()}; |
| 59 | |
| 60 | auto csrmv = |
| 61 | (is_csrmv_block ? common::getKernel("csrmv_thread", {{csrmv_cl_src}}, |
| 62 | targs, options) |
| 63 | : common::getKernel("csrmv_block", {{csrmv_cl_src}}, |
| 64 | targs, options)); |
| 65 | |
| 66 | int M = rowIdx.info.dims[0] - 1; |
| 67 | |
| 68 | int groups_x = |
| 69 | is_csrmv_block ? divup(M, REPEAT) : divup(M, REPEAT * local[0]); |
| 70 | groups_x = std::min(groups_x, MAX_CSRMV_GROUPS); |
| 71 | cl::NDRange global(local[0] * groups_x, 1); |
| 72 | |
| 73 | if (use_greedy) { |
| 74 | cl::Buffer *counter = bufferAlloc(sizeof(int)); |
| 75 | getQueue().enqueueFillBuffer(*counter, 0, 0, sizeof(int)); |
| 76 | csrmv(cl::EnqueueArgs(getQueue(), global, local), *out.data, |
| 77 | *values.data, *rowIdx.data, *colIdx.data, M, *rhs.data, rhs.info, |
| 78 | alpha, beta, *counter); |
| 79 | CL_DEBUG_FINISH(getQueue()); |
| 80 | bufferFree(counter); |
| 81 | } else { |
| 82 | csrmv(cl::EnqueueArgs(getQueue(), global, local), *out.data, |
| 83 | *values.data, *rowIdx.data, *colIdx.data, M, *rhs.data, rhs.info, |
| 84 | alpha, beta); |
| 85 | CL_DEBUG_FINISH(getQueue()); |
| 86 | } |
| 87 | } |
| 88 | } // namespace kernel |
no test coverage detected