| 26 | |
| 27 | template <typename T, typename Context> |
| 28 | void LogsumexpKernel(const Context& dev_ctx, |
| 29 | const DenseTensor& x, |
| 30 | const std::vector<int>& axis, |
| 31 | bool keepdim, |
| 32 | bool reduce_all, |
| 33 | DenseTensor* out) { |
| 34 | if (x.numel() == 0) { |
| 35 | Full<T, Context>(dev_ctx, out->dims(), -INFINITY, out); |
| 36 | return; |
| 37 | } |
| 38 | auto xdim = x.dims(); |
| 39 | for (int i = 0; i < xdim.size(); i++) |
| 40 | PADDLE_ENFORCE_LT(0, |
| 41 | xdim[i], |
| 42 | errors::InvalidArgument( |
| 43 | "The dims of Input(X) should be greater than 0.")); |
| 44 | |
| 45 | reduce_all = recompute_reduce_all(x, axis, reduce_all); |
| 46 | std::vector<int64_t> outdim_vec, keep_outdim_vec; |
| 47 | std::vector<int> axis_vec; |
| 48 | int64_t compute_size = 1, other_size = 1; |
| 49 | for (auto i : axis) { |
| 50 | auto v = i >= 0 ? i : i + xdim.size(); |
| 51 | axis_vec.push_back(v); |
| 52 | } |
| 53 | if (axis.size() == 0 || reduce_all) { |
| 54 | axis_vec.clear(); |
| 55 | for (int i = 0; i < xdim.size(); i++) { |
| 56 | axis_vec.push_back(i); |
| 57 | } |
| 58 | } |
| 59 | for (int i = 0; i < xdim.size(); i++) { |
| 60 | bool flag = false; |
| 61 | for (auto v : axis_vec) { |
| 62 | if (v == i) { |
| 63 | flag = true; |
| 64 | break; |
| 65 | } |
| 66 | } |
| 67 | if (flag) { |
| 68 | compute_size *= xdim[i]; |
| 69 | keep_outdim_vec.push_back(1); |
| 70 | if (keepdim) outdim_vec.push_back(1); |
| 71 | } else { |
| 72 | other_size *= xdim[i]; |
| 73 | outdim_vec.push_back(xdim[i]); |
| 74 | keep_outdim_vec.push_back(xdim[i]); |
| 75 | } |
| 76 | } |
| 77 | auto outdim = make_ddim(outdim_vec); |
| 78 | auto keep_outdim = make_ddim(keep_outdim_vec); |
| 79 | |
| 80 | // The XPU logsumexp api does not use xmax to normalize its input, so we |
| 81 | // fallback to the non fusion impl currently. |
| 82 | DenseTensor max_x; |
| 83 | max_x.Resize(keep_outdim); |
| 84 | MaxKernel<T, Context>(dev_ctx, x, axis_vec, true, &max_x); |
| 85 | |