vvchernov commented on code in PR #13894: URL: https://github.com/apache/tvm/pull/13894#discussion_r1104442490
########## python/tvm/topi/scatter_elements.py: ########## @@ -0,0 +1,150 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +"""ScatterElements operator""" +from tvm import te +from tvm import tir +from . import utils +from .math import cast + + +def scatter_elements(data, indices, updates, axis=0, reduction="update"): + """Scatter elements from updates to corresponding indices of copied data. + + Data, indices, updates and output have the same shape. + Indices can not have duplicates (if idx1 != idx2, then indices[idx1] != indices[idx2]) + if reduction == "update". + + .. code-block:: + + output[indices[i][j]][j] = f(output[indices[i][j]][j], updates[i][j]) if axis = 0 + output[i][indices[i][j]] = f(output[i][indices[i][j]], updates[i][j]) if axis = 1 + + where the update function f is determinted by the reduction. + Five types of the function are supported: "update", "add", "mul", "min" and "max" (see below) + + Parameters + ---------- + data : tvm.te.Tensor + The source array. + + indices : tvm.te.Tensor + The indices of the values to extract. + + updates : tvm.te.Tensor + The updates to apply at the Indices + + axis : optional, int + The axis to scatter on. It is zero by default. + + reduction : optional, string + The update mode for the algorithm, either "update", "add", "mul", "min" or "max" + If update, the update values will replace the input data + If add, the update values will be added to the input data + If mul, the update values will be multiply to the input data + If min, there is choice of minimal between the update values and the input data + If max, there is choice of maximal between the update values and the input data + It is "update" by default + + Returns + ------- + ret : tvm.te.Tensor + """ + if not isinstance(axis, int): + axis = utils.get_const_int(axis) + + # Prepare ranges and strides + shape = data.shape + if axis < 0: + axis = len(shape) + axis + axis_range = cast(shape[axis], indices.dtype) + + full_range = 1 + after_axis_range = 1 + for i, value in enumerate(shape, 0): + full_range *= value + if i > axis: + after_axis_range *= value + before_axis_stride = axis_range * after_axis_range + + ind_shape = indices.shape + ind_axis_range = ind_shape[axis] + + ind_before_axis_range = 1 + ind_after_axis_range = 1 + for i, value in enumerate(ind_shape, 0): + if i < axis: + ind_before_axis_range *= value + elif i > axis: + ind_after_axis_range *= value + ind_before_axis_stride = ind_axis_range * ind_after_axis_range + + def gen_ir(data_ptr, indices_ptr, updates_ptr, out_ptr): + # pylint: disable=invalid-name + ib = tir.ir_builder.create() + + data = ib.buffer_ptr(data_ptr) + indices = ib.buffer_ptr(indices_ptr) + updates = ib.buffer_ptr(updates_ptr) + out = ib.buffer_ptr(out_ptr) + + # Copy initial input data to output + with ib.for_range(0, full_range, "i", kind="parallel") as i: Review Comment: I tnink `parallel` is enough here. There are several reasons. 1) We can not use `vectorize` from the box: the loop should be tiled to get mostinner loop with size 4, 8, 16 or other specified degree of 2 (it is a question which one should be choosen). Another thing full_range is arbitrary, it is not limited by degre of 2, it means we should think additionally about loop tail. 2) This implementation is used by pure compilation. Of course, it does not give us the best performance, but it is assumed in TVM that the best performance can be gotten after metascheduling. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
