tmoreau89 commented on a change in pull request #4421: [VTA] Bringing group convolution support URL: https://github.com/apache/incubator-tvm/pull/4421#discussion_r351115576
########## File path: vta/tests/python/integration/test_benchmark_topi_group_conv2d.py ########## @@ -0,0 +1,242 @@ +# 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. + +"""Testing topi group conv2d operator for VTA""" + +import json +import os + +import numpy as np +from collections import namedtuple + +import tvm +from tvm import autotvm +from tvm.contrib import util +from tvm.contrib.pickle_memoize import memoize +import topi +import topi.testing +import vta +from vta import program_fpga, reconfig_runtime +import vta.testing +from vta.testing import simulator + + +Workload = namedtuple("GroupConv2DWorkload", + ['batch', 'height', 'width', 'in_filter', 'out_filter', 'groups', + 'hkernel', 'wkernel', 'hpad', 'wpad', 'hstride', 'wstride']) + +# Get batch info from env +env = vta.get_env() + +# Mobilenet (grouped variant) workloads +mobilenet_wkls = [ + ('mobilenet.D1', Workload(env.BATCH, 112, 112, 32, 32, 2, 3, 3, 1, 1, 1, 1)), + ('mobilenet.D2', Workload(env.BATCH, 112, 112, 64, 64, 4, 3, 3, 1, 1, 2, 2)), + ('mobilenet.D3', Workload(env.BATCH, 56, 56, 128, 128, 8, 3, 3, 1, 1, 1, 1)), + ('mobilenet.D4', Workload(env.BATCH, 56, 56, 128, 128, 8, 3, 3, 1, 1, 2, 2)), + ('mobilenet.D5', Workload(env.BATCH, 28, 28, 256, 256, 16, 3, 3, 1, 1, 1, 1)), + ('mobilenet.D6', Workload(env.BATCH, 28, 28, 256, 256, 16, 3, 3, 1, 1, 2, 2)), + ('mobilenet.D7', Workload(env.BATCH, 14, 14, 512, 512, 32, 3, 3, 1, 1, 1, 1)), + ('mobilenet.D8', Workload(env.BATCH, 14, 14, 512, 512, 32, 3, 3, 1, 1, 2, 2)), + ('mobilenet.D9', Workload(env.BATCH, 7, 7, 1024, 1024, 64, 3, 3, 1, 1, 1, 1)), +] + +# FIXME: we need a custom clip operator to circumvent a pattern detection limitation [email protected]_scope(tag=topi.tag.ELEMWISE) +def my_clip(x, a_min, a_max): + """Unlike topi's current clip, put min and max into two stages.""" + const_min = tvm.const(a_min, x.dtype) + const_max = tvm.const(a_max, x.dtype) + x = tvm.compute(x.shape, lambda *i: tvm.min(x(*i), const_max), name="clipA") + x = tvm.compute(x.shape, lambda *i: tvm.max(x(*i), const_min), name="clipB") + return x + +def run_group_conv2d(env, remote, wl, target, + check_correctness=True, print_ir=False, + samples=4): + + # Workload assertions + assert wl.hpad == wl.wpad + + # Perform packing only if we are targeting the accelerator + if "arm_cpu" in target.keys: + data_pack = False + layout = "NCHW" + elif "vta" in target.keys: + data_pack = True + layout = "NCHW%dn%dc" % (env.BATCH, env.BLOCK_IN) + + # Derive shapes depending upon packing + CI_G = wl.in_filter // wl.groups + a_shape = (wl.batch, wl.in_filter, wl.height, wl.width) + w_shape = (wl.out_filter, CI_G, wl.hkernel, wl.wkernel) + b_shape = (wl.batch, wl.out_filter, 1, 1) + if data_pack: + data_shape = (wl.batch//env.BATCH, wl.in_filter//env.BLOCK_IN, + wl.height, wl.width, env.BATCH, env.BLOCK_IN) + kernel_shape = (wl.out_filter//env.BLOCK_OUT, CI_G//env.BLOCK_IN, + wl.hkernel, wl.wkernel, env.BLOCK_OUT, env.BLOCK_IN) + bias_shape = (wl.batch//env.BATCH, wl.out_filter//env.BLOCK_OUT, + 1, 1, env.BATCH, env.BLOCK_OUT) + else: + data_shape = a_shape + kernel_shape = w_shape + bias_shape = b_shape + data = tvm.placeholder(data_shape, name="data", dtype=env.inp_dtype) + kernel = tvm.placeholder(kernel_shape, name="kernel", dtype=env.wgt_dtype) + bias = tvm.placeholder(bias_shape, name="bias", dtype=env.acc_dtype) + # Define base computation schedule + with target: + res = topi.nn.group_conv2d_nchw( + data, kernel, (wl.hstride, wl.wstride), (wl.hpad, wl.wpad), (1, 1), + wl.groups, env.acc_dtype) + res = topi.right_shift(res, 8) + res = topi.add(res, bias) + res = my_clip(res, 0, (1 << env.OUT_WIDTH - 1) - 1) + res = topi.cast(res, env.out_dtype) + # Derive base schedule + s = topi.generic.schedule_group_conv2d_nchw([res]) + if print_ir: + print(vta.lower(s, [data, kernel, bias, res], simple_mode=True)) + + # Derive number of ops + fout_height = (wl.height + 2 * wl.hpad - wl.hkernel) // wl.hstride + 1 + fout_width = (wl.width + 2 * wl.wpad - wl.wkernel) // wl.wstride + 1 + num_ops = 2 * wl.batch * fout_height * fout_width * wl.hkernel * wl.wkernel * \ + wl.out_filter * wl.in_filter // wl.groups + + # @memoize("vta.tests.test_benchmark_topi.conv2d.verify_nhwc") Review comment: We can remove it ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
