mbaret commented on a change in pull request #8795: URL: https://github.com/apache/tvm/pull/8795#discussion_r700175964
########## File path: python/tvm/relay/backend/contrib/ethosu/te/convolution.py ########## @@ -0,0 +1,199 @@ +# 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-argument +"""Tensor Expressions for convolutions for the NPU""" +from tvm import te +from .dma import dma_ofm_compute, dma_ifm_compute + + +def process_stride(stride): + """Process the striding into a common format. + + Parameters + ---------- + stride : Union[int, tuple, list] + The 2D striding. + int -> striding is the same in the height and width axis. + 2D -> striding specified as (stride height, stride width). + + Returns + ------- + int + The stride in the height axis. + int + The stride in the width axis. + + """ + assert isinstance(stride, int) or len(stride) == 2 + if isinstance(stride, int): + return stride, stride + + return stride + + +def process_dilation(dilation): + """Process the dilation into a common format. + + Parameters + ---------- + dilation : Union[int, tuple, list] + The 2D dilation. + int -> dilation is the same in the height and width axis. + 2D -> dilation specified as (dilation height, dilation width). + + Returns + ------- + int + The dilation in the height axis. + int + The dilation in the width axis. + + """ + assert isinstance(dilation, int) or len(dilation) == 2 + if isinstance(dilation, int): + return dilation, dilation + + return dilation + + +def conv2d_compute( + ifm, + weight, + scale_bias, + lut, + ifm_scale, + ifm_zero_point, + weight_zero_point, + ofm_scale, + ofm_zero_point, + strides, + padding, + dilation, + activation, + clip_min, + clip_max, + upscale, + ifm_layout, + ofm_layout, +): + """A compute operator representing the capabilities of a 2D convolution for the NPU. + + Parameters + ---------- + ifm : te.Tensor + The Input Feature Map tensor (IFM). + weight : te.Tensor + The weight tensor. + scale_bias : te.Tensor + The packed per-channel weight scale and bias tensor. + lut : te.Tensor + The look-up table values to use if activation = "LUT". + ifm_scale : float + The quantization scale for the Input Feature Map tensor. + ifm_zero_point : int + The quantization zero point for the Input Feature Map tensor. + weight_zero_point : int + The quantization zero point for the weight tensor. + ofm_scale : float + The quantization scale for the Output Feature Map tensor. + ofm_zero_point : int + The quantization zero point for the Output Feature Map tensor. + strides : tuple + The 2 dimensional strides as (stride_height, stride_width). + padding : tuple + The 4 dimensional padding as (pad_top, pad_left, pad_bottom, pad_right). + dilation : Union[int, tuple, list] + The 2 dimensional dilation as (dilation_height, dilation_width). + activation : str + The activation function to use. + "NONE" - no activation function. + "CLIP" - clip the output between clip_min and clip_max. + "TANH" - tanh activation function. + "SIGMOID" - sigmoid activation function. + "LUT" - use a look-up table to perform the activation function. + clip_min : int + The minimum clipping value if activation = "CLIP". + clip_max : int + The maximum clipping value if activation = "CLIP". + upscale : str + The 2x2 upscaling mode to apply to the Input Feature Map tensor. + "NONE" - no upscaling. + "NEAREST" - upscale using nearest neighbour. + "ZEROS" - upscale using zeros. + ifm_layout : str + The layout of the Input Feature Map tensor. Can be "NHWC" or "NHCWB16". + ofm_layout : str + The layout of the Output Feature Map tensor. Can be "NHWC" or "NHCWB16". + + Returns + ------- + te.Tensor + The OFM tensor. + + """ + assert ifm.shape[0] == 1 + assert ifm_layout in {"NHWC", "NHCWB16"} + assert ofm_layout in {"NHWC", "NHCWB16"} + + stride_h, stride_w = strides + dilation_h, dilation_w = dilation + ofm_channels, kernel_h, kernel_w, ifm_channels = weight.shape + + # Compute operation for the IFM DMA pipeline + dmaed_ifm = dma_ifm_compute( + ifm, ifm_layout, ifm_zero_point, ifm_scale, weight.shape[3], padding + ) + + # 2D Convolution compute operation + dilated_kernel_h = (kernel_h - 1) * dilation_h + 1 + dilated_kernel_w = (kernel_w - 1) * dilation_w + 1 Review comment: I think we can address that when we introduce another operator that has similar logic. That way we can see if it's worth pulling out and if so what the correct interface should be. -- 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]
