manupa-arm commented on a change in pull request #9209: URL: https://github.com/apache/tvm/pull/9209#discussion_r723866285
########## File path: python/tvm/relay/backend/contrib/ethosu/tir/depthwise.py ########## @@ -0,0 +1,111 @@ +# 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 +"""Extract information from the depthwise convolution operators in TIR.""" +import tvm +from ..vela_api import SCALE_BIAS_LENGTH +from .utils import get_outer_loops, get_op_attrs, get_base_address, get_loads, get_stores +from .dma import get_ifm_params, get_ofm_params +from .spec import ( + SerialKernel, + SerialAddressRange, + SerialActivation, + Serial2DDepthwise, +) + + +def get_depthwise2d_params(stmt, producers, consumers): Review comment: nit : type annotations ########## File path: python/tvm/relay/backend/contrib/ethosu/legalize.py ########## @@ -208,6 +208,96 @@ def __call__(self, *args, **kwargs): pass +class EthosuDepthwise2DRewriter(DFPatternCallback): + """Convert ethosu.qnn_depthwise2d composite functions to ethosu_depthwise2d operators""" Review comment: Let us stick to depthwiseconv2d/DepthwiseConv2D and also in the following mentions to it. ########## File path: python/tvm/relay/backend/contrib/ethosu/te/depthwise.py ########## @@ -0,0 +1,148 @@ +# 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 depthwise convolutions""" +from typing import Tuple, Union, List + +from tvm import te +from .dma import dma_ofm_compute, dma_ifm_compute + + +def depthwise2d_compute( + ifm: te.Tensor, + weight: te.Tensor, + scale_bias: te.Tensor, + lut: te.Tensor, + ifm_scale: float, + ifm_zero_point: int, + weight_zero_point: int, + ofm_scale: float, + ofm_zero_point: int, + strides: Tuple[int, int], + padding: Tuple[int, int, int, int], + dilation: Union[Tuple[int, int], List[int]], + activation: str, + clip_min: int, + clip_max: int, + upscale: str, + ifm_layout: str, + ofm_layout: str, +) -> te.Tensor: + """A compute operator representing the capabilities of 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 Review comment: It is better to give a message when this fails as to why it was assumed to be 1. ########## File path: python/tvm/relay/backend/contrib/ethosu/op/depthwise.py ########## @@ -0,0 +1,205 @@ +# 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=unused-argument +"""Relay operator for depthwise convolution""" +from typing import Tuple + +import tvm +from tvm.relay.op import _make +from tvm.topi.generic import schedule_injective +from tvm.relay.op.op import OpStrategy +from tvm.relay.op import strategy as _strategy + +from ..te import depthwise2d_compute + + +def _extract_ethosu_depthwise2d_params(attrs, args): + """Get the parameters necessary to construct a ethosu_depthwise2d compute TE + from a ethosu_depthwise2d Relay call.""" + ifm = args[0] + weight = args[1] + scale_bias = args[2] + lut = args[3] + ifm_scale = attrs.ifm_scale + ifm_zero_point = attrs.ifm_zero_point + weight_zero_point = attrs.weight_zero_point + ofm_scale = attrs.ofm_scale + ofm_zero_point = attrs.ofm_zero_point + strides = attrs.strides + padding = attrs.padding + dilation = attrs.dilation + activation = attrs.activation + clip_min = attrs.clip_min + clip_max = attrs.clip_max + upscale = attrs.upscale + ifm_layout = attrs.ifm_layout + ofm_layout = attrs.ofm_layout + + return ( + 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, + ) + + [email protected]_op_attr("contrib.ethosu.depthwise2d", "FTVMCompute") +def create_ethosu_depthwise2d_compute(attrs, args, out_type): + """Create an ethosu_depthwise2d compute op.""" + params = _extract_ethosu_depthwise2d_params(attrs, args) + op = depthwise2d_compute(*params) + return [op] + + [email protected]_op_attr("contrib.ethosu.depthwise2d", "FTVMStrategy") +def depthwise2d_strategy_ethosu(attrs, inputs, out_type, target): + strategy = OpStrategy() + strategy.add_implementation( + create_ethosu_depthwise2d_compute, + _strategy.wrap_topi_schedule(schedule_injective), + name="ethosu_depthwise2d", + ) + return strategy + + +def ethosu_depthwise2d( + ifm: tvm.relay.Expr, + weight: tvm.relay.Expr, + scale_bias: tvm.relay.Expr, + lut: tvm.relay.Expr, + ifm_scale: float, + ifm_zero_point: int, + weight_zero_point: int, + ofm_scale: float, + ofm_zero_point: int, + kernel_shape: Tuple[int, int], + ofm_channels: int, + strides: Tuple[int, int] = (1, 1), Review comment: nit : We can use Optional[Tuple[int, int]] ########## File path: python/tvm/relay/backend/contrib/ethosu/legalize.py ########## @@ -208,6 +208,96 @@ def __call__(self, *args, **kwargs): pass +class EthosuDepthwise2DRewriter(DFPatternCallback): + """Convert ethosu.qnn_depthwise2d composite functions to ethosu_depthwise2d operators""" + + def __init__(self): + super().__init__(require_type=True) + self.pattern = ( + wildcard().has_attr({"Composite": ethosu_patterns.QnnDepthwise2DParams.composite_name}) Review comment: QnnDepthwiseConv2DParams ########## File path: tests/python/contrib/test_ethosu/test_legalize.py ########## @@ -20,15 +20,29 @@ pytest.importorskip("ethosu.vela") import numpy as np +import tensorflow as tf import tvm from tvm import relay -from tvm.relay.backend.contrib import ethosu from tvm.relay.backend.contrib.ethosu import legalize, preprocess -from tvm.relay.dataflow_pattern import * -from tvm.relay.op.contrib.ethosu import * +from tvm.relay import dataflow_pattern +from tvm.relay.op.contrib import ethosu +from tvm.relay.build_module import bind_params_by_name from . import relay_ir_builder +from . import infra + + +def partition_ethosu_by_table(mod, pattern_table): Review comment: Maybe add a docstring to explain why we have this method as opposed to the standard partition function. -- 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]
