lhutton1 commented on code in PR #12353: URL: https://github.com/apache/tvm/pull/12353#discussion_r948992261
########## tests/python/contrib/test_cmsisnn/test_fuse_pads.py: ########## @@ -0,0 +1,340 @@ +# 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. + +"""CMSIS-NN integration tests: fuse_pads pass""" +import numpy as np +import pytest +import tvm +import tvm.testing +from tvm import relay +from .utils import CheckForPadsWithinCompositeFunc + +tvm._ffi._init_api("relay.ext.cmsisnn.transform", __name__) + + +def set_external_func_attr(func, compiler, ext_symbol): + func = func.with_attr("Primitive", tvm.tir.IntImm("int32", 1)) + func = func.with_attr("Compiler", compiler) + func = func.with_attr("global_symbol", ext_symbol) + return func + + +def set_composite_func_attr(func, name): + func = func.with_attr("Composite", name) + return func + + [email protected]( + "ifm_shape, pad_width, conv2d_padding, ofm_shape", + [ + [(1, 25, 25, 12), ((0, 0), (0, 2), (1, 2), (0, 0)), (1, 1, 1, 1), (1, 26, 28, 2)], + [(1, 64, 100, 4), ((0, 0), (1, 3), (1, 1), (0, 0)), (0, 0, 0, 0), (1, 64, 100, 2)], + [(1, 55, 55, 3), ((0, 0), (2, 1), (3, 5), (0, 0)), (0, 0, 1, 1), (1, 57, 59, 2)], + ], +) +def test_invalid_padding_for_fusion(ifm_shape, pad_width, conv2d_padding, ofm_shape): + """Negative tests for pads preceding Conv2D that cannot be fused.""" + dtype = "int8" + kernel_size = (3, 3) + ofm_channels = 2 + local_input = relay.var("local_input", shape=ifm_shape, dtype=dtype) + pad = relay.nn.pad( + local_input, + pad_width=pad_width, # ((), (top, bottom), (left, right), ()) + pad_value=10, + pad_mode="constant", + ) + rng = np.random.default_rng(12321) + local_weight = tvm.nd.array( + rng.integers( + np.iinfo(dtype).min, + high=np.iinfo(dtype).max, + size=(ofm_channels, kernel_size[0], kernel_size[1], ifm_shape[3]), + dtype=dtype, + ) + ) + local_weight = relay.const(local_weight, dtype) + conv2d = relay.qnn.op.conv2d( + pad, + local_weight, + relay.const(1, "int32"), + relay.const(1, "int32"), + relay.const(1, "float32"), + relay.const(1, "float32"), + data_layout="NHWC", + kernel_layout="OHWI", + channels=ofm_channels, + kernel_size=(3, 3), + padding=conv2d_padding, + out_dtype="int32", + ) + requantize = relay.qnn.op.requantize( + conv2d, + relay.const(1, "float32"), + relay.const(1, "int32"), + relay.const(1, "float32"), + relay.const(1, "int32"), + axis=0, + out_dtype=dtype, + ) + local_func = relay.Function(relay.analysis.free_vars(requantize), requantize) + local_func = set_composite_func_attr(local_func, "cmsis-nn.qnn_conv2d") + + mod = tvm.IRModule() + ext_input = relay.var("ext_input", shape=ifm_shape, dtype=dtype) + call_local_func = relay.Call(local_func, [ext_input]) + extern_func = relay.Function(relay.analysis.free_vars(call_local_func), call_local_func) + extern_var = relay.GlobalVar("external_function") + extern_func = set_external_func_attr(extern_func, "cmsis-nn", extern_var.name_hint) + mod[extern_var] = extern_func + + main_input = relay.var("main_input", shape=ifm_shape, dtype=dtype) + call_extern_func = relay.Call(extern_var, [main_input]) + main_func = relay.Function([main_input], call_extern_func, relay.TensorType(ofm_shape, dtype)) + main_var = relay.GlobalVar("main") + mod[main_var] = main_func + + mod = relay.transform.InferType()(mod) + + error_regex = r"Difference on each side of a dimension should be either 0 or 1" Review Comment: Nit: possible to check the whole error message? i.e. with the pad values? ########## python/tvm/relay/op/contrib/cmsisnn.py: ########## @@ -136,15 +145,72 @@ def check_qnn_conv2d(pattern): ): is_depthwise = True - return ( + ret = ( + conv2d.attrs.out_dtype == "int32" + and conv2d_input.checked_type.dtype == "int8" + and conv2d_weight.checked_type.dtype == "int8" + and pattern.checked_type.dtype == "int8" + and bias_dtype == "int32" + and all([zp == 0 for zp in kernel_zp]) + and (not is_depthwise or bias_add is not None) + ) + return ret + + def check_qnn_conv2d_pad(pattern): Review Comment: Since some of the logic here is shared between `check_qnn_conv2d` and `check_qnn_conv2d_pad`, could we create a generic function that returns the appropriate check function? e.g. `create_check_qnn_conv2d(with_pad=False)` and then in the pattern table we can have something like: ``` ("cmsis-nn.qnn_conv2d", qnn_conv2d_pattern(with_pad=True), create_check_conv2d(True)), ("cmsis-nn.qnn_conv2d", qnn_conv2d_pattern(with_pad=False), create_check_conv2d(False)), ``` ########## src/relay/backend/contrib/cmsisnn/fuse_pads.cc: ########## @@ -0,0 +1,219 @@ + +/* + * 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. + */ +/*! + * \file fuse_pads.cc + * \brief Fuses pads that precede qnn.conv2d ops inside CMSIS-NN composite functions. + */ + +#include <tvm/relay/attrs/nn.h> +#include <tvm/relay/attrs/transform.h> +#include <tvm/relay/expr_functor.h> +#include <tvm/relay/transform.h> +#include <tvm/runtime/ndarray.h> + +#include "../../../op/make_op.h" +#include "../../../qnn/utils.h" +#include "../../../transforms/pattern_utils.h" +#include "convolutions.h" + +namespace tvm { +namespace relay { +namespace contrib { +namespace cmsisnn { + +/*! + * \brief This Mutator will find all partitioned functions meant for CMSIS-NN Conv2D. + * Then, it will fuse preceding pads with qnn.conv2d. + */ +class FusePadsMutator : public MixedModeMutator { + public: + explicit FusePadsMutator(const IRModule& mod) : mod_(mod) {} + + private: + /*! * \brief In order to eliminate preceding nn.pad op, pad_width of nn.pad is passed onto + * convolution layer to update Conv2DAttrs's padding attribute. */ + void UpdateConv2DPadding(const CallNode* conv2d_call, const Array<Array<Integer>>& pad_width, + const Conv2DAttrs* conv2d_attrs, Attrs* new_attrs) { + auto attrs = make_object<Conv2DAttrs>(); + attrs->strides = std::move(conv2d_attrs->strides); + attrs->dilation = std::move(conv2d_attrs->dilation); + attrs->groups = conv2d_attrs->groups; + attrs->channels = std::move(conv2d_attrs->channels); + attrs->kernel_size = std::move(conv2d_attrs->kernel_size); + attrs->data_layout = std::move(conv2d_attrs->data_layout); + attrs->kernel_layout = std::move(conv2d_attrs->kernel_layout); + attrs->out_layout = std::move(conv2d_attrs->out_layout); + attrs->out_dtype = std::move(conv2d_attrs->out_dtype); + + // pad_width: ((), (top, bottom), (left, right), ()) for NHWC layout + // conv2d_attrs->padding: (top, left, bottom, right) + std::string data_layout = conv2d_attrs->data_layout.c_str(); + int pos_h = data_layout.find("H"); + int pos_w = data_layout.find("W"); + + int pad_top = + qnn::get_const_int(conv2d_attrs->padding[0]) + qnn::get_const_int(pad_width[pos_h][0]); + int pad_left = + qnn::get_const_int(conv2d_attrs->padding[1]) + qnn::get_const_int(pad_width[pos_w][0]); + int pad_bottom = + qnn::get_const_int(conv2d_attrs->padding[2]) + qnn::get_const_int(pad_width[pos_h][1]); + int pad_right = + qnn::get_const_int(conv2d_attrs->padding[3]) + qnn::get_const_int(pad_width[pos_w][1]); + + int pad_diff_w = pad_right - pad_left; + int pad_diff_h = pad_bottom - pad_top; Review Comment: Nice :) ########## src/relay/backend/contrib/cmsisnn/fuse_pads.cc: ########## @@ -0,0 +1,208 @@ +/* + * 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. + */ +/*! + * \file src/relay/backend/contrib/cmsisnn/fuse_pads.cc + * \brief Fuses pads that precede qnn.conv2d ops inside CMSIS-NN composite functions. + */ + +#include <tvm/relay/attrs/nn.h> +#include <tvm/relay/attrs/transform.h> +#include <tvm/relay/expr_functor.h> +#include <tvm/relay/transform.h> +#include <tvm/runtime/ndarray.h> + +#include "../../../op/make_op.h" +#include "../../../qnn/utils.h" +#include "../../../transforms/pattern_utils.h" +#include "convolutions.h" + +namespace tvm { +namespace relay { +namespace contrib { +namespace cmsisnn { + +inline IntImm ToIntImm(int32_t value) { return IntImm(DataType::Int(32), value); } + +/*! + * \brief From padding attributes of nn.pad and qnn.conv2d, calculates effective padding along H + * and W dimensions. + */ +Array<IntImm> GetEffectiveConv2DPadding(Expr conv2d, Expr pad) { + // pad_width: ((), (top, bottom), (left, right), ()) for NHWC layout + // conv2d_attrs->padding: (top, left, bottom, right) + auto* conv2d_call = conv2d.as<CallNode>(); + auto* conv2d_attrs = conv2d_call->attrs.as<Conv2DAttrs>(); + std::string data_layout = conv2d_attrs->data_layout.c_str(); + int pos_h = data_layout.find("H"); + int pos_w = data_layout.find("W"); + + auto* pad_call = pad.as<CallNode>(); + Array<Array<Integer>> pad_width = pad_call->attrs.as<PadAttrs>()->pad_width; + int pad_top = + qnn::get_const_int(conv2d_attrs->padding[0]) + qnn::get_const_int(pad_width[pos_h][0]); + int pad_left = + qnn::get_const_int(conv2d_attrs->padding[1]) + qnn::get_const_int(pad_width[pos_w][0]); + int pad_bottom = + qnn::get_const_int(conv2d_attrs->padding[2]) + qnn::get_const_int(pad_width[pos_h][1]); + int pad_right = + qnn::get_const_int(conv2d_attrs->padding[3]) + qnn::get_const_int(pad_width[pos_w][1]); + + return {ToIntImm(pad_top), ToIntImm(pad_left), ToIntImm(pad_bottom), ToIntImm(pad_right)}; +} + +/*! + * \brief This Mutator will find all partitioned functions meant for CMSIS-NN Conv2D. + * Then, it will fuse preceding pads with qnn.conv2d. + */ +class FusePadsMutator : public MixedModeMutator { + public: + explicit FusePadsMutator(const IRModule& mod) : mod_(mod) {} + + private: + /*! * \brief In order to eliminate preceding nn.pad op, pad_width of nn.pad is passed onto Review Comment: Nit: newline after "!" -- 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]
