| 108 | |
| 109 | template<typename Ti, typename To, af_op_t op> |
| 110 | static void scanFirst(Param &out, const Param &in, |
| 111 | const bool inclusiveScan = true) { |
| 112 | uint threads_x = nextpow2(std::max(32u, (uint)out.info.dims[0])); |
| 113 | threads_x = std::min(threads_x, THREADS_PER_GROUP); |
| 114 | uint threads_y = THREADS_PER_GROUP / threads_x; |
| 115 | |
| 116 | uint groups_x = divup(out.info.dims[0], threads_x * REPEAT); |
| 117 | uint groups_y = divup(out.info.dims[1], threads_y); |
| 118 | |
| 119 | if (groups_x == 1) { |
| 120 | scanFirstLauncher<Ti, To, op>(out, out, in, true, groups_x, groups_y, |
| 121 | threads_x, inclusiveScan); |
| 122 | |
| 123 | } else { |
| 124 | Param tmp = out; |
| 125 | tmp.info.dims[0] = groups_x; |
| 126 | tmp.info.strides[0] = 1; |
| 127 | for (int k = 1; k < 4; k++) { |
| 128 | tmp.info.strides[k] = |
| 129 | tmp.info.strides[k - 1] * tmp.info.dims[k - 1]; |
| 130 | } |
| 131 | |
| 132 | int tmp_elements = tmp.info.strides[3] * tmp.info.dims[3]; |
| 133 | |
| 134 | tmp.data = bufferAlloc(tmp_elements * sizeof(To)); |
| 135 | |
| 136 | scanFirstLauncher<Ti, To, op>(out, tmp, in, false, groups_x, groups_y, |
| 137 | threads_x, inclusiveScan); |
| 138 | |
| 139 | if (op == af_notzero_t) { |
| 140 | scanFirstLauncher<To, To, af_add_t>(tmp, tmp, tmp, true, 1, |
| 141 | groups_y, threads_x, true); |
| 142 | } else { |
| 143 | scanFirstLauncher<To, To, op>(tmp, tmp, tmp, true, 1, groups_y, |
| 144 | threads_x, true); |
| 145 | } |
| 146 | |
| 147 | bcastFirstLauncher<To, To, op>(out, tmp, true, groups_x, groups_y, |
| 148 | threads_x, inclusiveScan); |
| 149 | |
| 150 | bufferFree(tmp.data); |
| 151 | } |
| 152 | } |
| 153 | |
| 154 | } // namespace kernel |
| 155 | } // namespace opencl |
nothing calls this directly
no test coverage detected