echuraev commented on code in PR #11543: URL: https://github.com/apache/tvm/pull/11543#discussion_r889262027
########## python/tvm/topi/adreno/conv2d_nchw_winograd.py: ########## @@ -0,0 +1,481 @@ +# 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. +# pylint: disable=invalid-name,unused-variable,unused-argument +"""Winograd template for Adreno backend""" + +import logging +import tvm +from tvm import te +from tvm import autotvm + +from tvm.topi import nn +from tvm.topi.utils import get_const_int, get_const_tuple, traverse_inline +from ..nn.winograd_util import winograd_transform_matrices +from .utils import ( + split_to_chunks, + pack_input, + pack_filter, + bind_data_copy, + get_texture_storage, +) + + +logger = logging.getLogger("conv2d_nchw_winograd") + + +def _infer_tile_size(data): + if len(data.shape) == 4: + N, CI, H, W = get_const_tuple(data.shape) + else: + N, CI, H, W, CB = get_const_tuple(data.shape) + + if H % 8 == 0: + return 4 + return 2 + + [email protected]_topi_compute("conv2d_nchw_winograd.image2d") +def conv2d_nchw_winograd(cfg, data, kernel, strides, padding, dilation, out_dtype): + args = {"shared": False, "accumulator": "float16"} + return conv2d_nchw_winograd_comp( + cfg, data, kernel, strides, padding, dilation, out_dtype, args=args, pre_computed=False + ) + + [email protected]_topi_compute("conv2d_nchw_winograd_acc32.image2d") +def conv2d_nchw_winograd_acc32(cfg, data, kernel, strides, padding, dilation, out_dtype): + args = {"shared": False, "accumulator": "float32"} + return conv2d_nchw_winograd_comp( + cfg, data, kernel, strides, padding, dilation, out_dtype, args=args, pre_computed=False + ) + + [email protected]_topi_schedule("conv2d_nchw_winograd.image2d") +def schedule_conv2d_nchw_winograd(cfg, outs): + return schedule_conv2d_nchw_winograd_impl(cfg, outs, tag="cast_from_acc16") + + [email protected]_topi_schedule("conv2d_nchw_winograd_acc32.image2d") +def schedule_conv2d_nchw_winograd_acc32(cfg, outs): + return schedule_conv2d_nchw_winograd_impl(cfg, outs, tag="cast_from_acc32") + + [email protected]_topi_compute("conv2d_nchw_winograd_without_weight_transform.image2d") +def conv2d_nchw_winograd_without_weight_transform( + cfg, data, kernel, strides, padding, dilation, out_dtype +): + args = {"shared": False, "accumulator": "float16"} + return conv2d_nchw_winograd_comp( + cfg, data, kernel, strides, padding, dilation, out_dtype, args=args, pre_computed=True + ) + + [email protected]_topi_compute("conv2d_nchw_winograd_without_weight_transform_acc32.image2d") +def conv2d_nchw_winograd_without_weight_transform_acc32( + cfg, data, kernel, strides, padding, dilation, out_dtype +): + args = {"shared": False, "accumulator": "float32"} + return conv2d_nchw_winograd_comp( + cfg, data, kernel, strides, padding, dilation, out_dtype, args=args, pre_computed=True + ) + + [email protected]_topi_schedule("conv2d_nchw_winograd_without_weight_transform.image2d") +def schedule_conv2d_nchw_winograd_without_weight_transform(cfg, outs): + return schedule_conv2d_nchw_winograd_impl(cfg, outs, tag="cast_from_acc16", pre_computed=True) + + [email protected]_topi_schedule("conv2d_nchw_winograd_without_weight_transform_acc32.image2d") +def schedule_conv2d_nchw_winograd_without_weight_transform_acc32(cfg, outs): + return schedule_conv2d_nchw_winograd_impl(cfg, outs, tag="cast_from_acc32", pre_computed=True) + + +def schedule_conv2d_nchw_winograd_impl(cfg, outs, tag, pre_computed=False): + outs = [outs] if isinstance(outs, te.tensor.Tensor) else outs + s = te.create_schedule([x.op for x in outs]) + + def _callback(op): + if op.tag == tag: + schedule_conv2d_winograd(cfg, s, op.output(0), pre_computed=pre_computed) + + traverse_inline(s, outs[0].op, _callback) + return s + + +def conv2d_nchw_winograd_comp( + cfg, data, kernel, strides, padding, dilation, out_dtype, args, pre_computed Review Comment: 1. `pre_computed` means that the weights were precomputed and we should do any transformation with them. 2. I'll add a documentation -- 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]
