1# Licensed to the Apache Software Foundation (ASF) under one 2# or more contributor license agreements. See the NOTICE file 3# distributed with this work for additional information 4# regarding copyright ownership. The ASF licenses this file 5# to you under the Apache License, Version 2.0 (the 6# "License"); you may not use this file except in compliance 7# with the License. You may obtain a copy of the License at 8# 9# http://www.apache.org/licenses/LICENSE-2.0 10# 11# Unless required by applicable law or agreed to in writing, 12# software distributed under the License is distributed on an 13# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY 14# KIND, either express or implied. See the License for the 15# specific language governing permissions and limitations 16# under the License. 17 18import tvm 19import numpy as np 20from tvm import relay 21from tvm.contrib import graph_runtime 22 23roundings = ["UPWARD", "TONEAREST"] 24 25def verify(mod, goldens): 26 with relay.build_config(opt_level=3): 27 graph, lib, params = relay.build(mod, "llvm", params=None) 28 golden_data, golden_output = goldens 29 rt_mod = graph_runtime.create(graph, lib, ctx=tvm.cpu(0)) 30 rt_mod.set_input("quantized_data",golden_data) 31 rt_mod.set_input(**params) 32 rt_mod.run() 33 res = rt_mod.get_output(0).asnumpy() 34 np.testing.assert_equal(res, golden_output) 35 36def get_mod(data_shape, data_dtype, out_dtype, input_scale, output_scale, 37 input_zero_point=0, output_zero_point=0, rounding="TONEAREST"): 38 quantized_data = relay.var("quantized_data", shape=data_shape, 39 dtype=data_dtype) 40 mod = relay.qnn.op.requantize( 41 quantized_data, 42 input_scale=input_scale, 43 input_zero_point=input_zero_point, 44 output_scale=output_scale, 45 output_zero_point=output_zero_point, 46 rounding=rounding, 47 out_dtype=out_dtype) 48 49 mod = relay.Function(relay.analysis.free_vars(mod), mod) 50 mod = relay.Module.from_expr(mod) 51 return mod 52 53def test_same_scale(): 54 # Have same scales, everything within range 55 golden_data = np.arange(-100, 100, 1).astype('int32') 56 golden_output = golden_data 57 58 for rounding in roundings: 59 mod = get_mod(data_shape=(200, ), 60 data_dtype='int32', 61 out_dtype="int8", 62 input_scale=0.5, 63 output_scale=0.5, 64 rounding=rounding) 65 assert 'right_shift' not in mod.astext() 66 verify(mod, (golden_data, golden_output)) 67 68def test_downscale(): 69 for rounding in roundings: 70 mod = get_mod(data_shape=(32, ), 71 data_dtype='int32', 72 out_dtype='int8', 73 input_scale=1, 74 output_scale=16, 75 rounding=rounding) 76 77 # Try positive values 78 # 8 corresponds to 0.5, resulting in 1 79 golden_data = np.arange(0, 32, 1).astype('int32') 80 golden_output = np.repeat([0, 1, 2], [8, 16, 8]) 81 verify(mod, (golden_data, golden_output)) 82 83 # Try negative values 84 # -8 corresponds to -0.5. For UPWARD, this is 0 85 golden_data = np.arange(0, -32, -1).astype('int32') 86 if rounding == "UPWARD": 87 golden_output = np.repeat([0, -1, -2], [9, 16, 7]) 88 else: 89 golden_output = np.repeat([0, -1, -2], [8, 16, 8]) 90 verify(mod, (golden_data, golden_output)) 91 92 # Try a different scale 93 mod = get_mod(data_shape=(32, ), 94 data_dtype='int32', 95 out_dtype="int8", 96 input_scale=1, 97 output_scale=4, 98 rounding=rounding) 99 100 # Try positive values 101 # 2I corresponds to 0.5, resulting in 1 102 golden_data = np.arange(0, 32, 1).astype('int32') 103 golden_output = np.repeat([0, 1, 2, 3, 4, 5, 6, 7, 8], 104 [2, 4, 4, 4, 4, 4, 4, 4, 2]) 105 verify(mod, (golden_data, golden_output)) 106 107 # Try negative values 108 # -8 corresponds to -0.5. For UPWARD, this is 0 109 golden_data = np.arange(0, -32, -1).astype('int32') 110 if rounding == "UPWARD": 111 golden_output = np.repeat([0, -1, -2, -3, -4, -5, -6, -7, -8], 112 [3, 4, 4, 4, 4, 4, 4, 4, 1]) 113 else: 114 golden_output = np.repeat([0, -1, -2, -3, -4, -5, -6, -7, -8], 115 [2, 4, 4, 4, 4, 4, 4, 4, 2]) 116 verify(mod, (golden_data, golden_output)) 117 118 # Try uint8 out_dtype 119 mod = get_mod(data_shape=(32, ), 120 data_dtype='int32', 121 out_dtype='uint8', 122 input_scale=1, 123 output_scale=16, 124 rounding=rounding) 125 126 # Try positive values 127 # 8 corresponds to 0.5, resulting in 1 128 golden_data = np.arange(0, 32, 1).astype('int32') 129 golden_output = np.repeat([0, 1, 2], [8, 16, 8]) 130 verify(mod, (golden_data, golden_output)) 131 132 # Try uint8 in_dtyope and uint8 out_dtype 133 mod = get_mod(data_shape=(32, ), 134 data_dtype='uint8', 135 out_dtype='uint8', 136 input_scale=1, 137 output_scale=16, 138 rounding=rounding) 139 140 # Try positive values 141 # 8 corresponds to 0.5, resulting in 1 142 golden_data = np.arange(0, 32, 1).astype('int32') 143 golden_output = np.repeat([0, 1, 2], [8, 16, 8]) 144 verify(mod, (golden_data, golden_output)) 145 146def test_upscale(): 147 for rounding in roundings: 148 mod = get_mod(data_shape=(32, ), 149 data_dtype='int32', 150 out_dtype="int8", 151 input_scale=2, 152 output_scale=1, 153 rounding=rounding) 154 155 # Try positive values 156 # 8 corresponds to 0.5, resulting in 1 157 golden_data = np.arange(0, 32, 1).astype('int32') 158 golden_output = np.multiply(2, golden_data) 159 verify(mod, (golden_data, golden_output)) 160 161 # Try negative values 162 # -8 corresponds to -0.5. For UPWARD, this is 0 163 golden_data = np.arange(0, -32, -1).astype('int32') 164 golden_output = np.multiply(2, golden_data) 165 verify(mod, (golden_data, golden_output)) 166 167def test_saturation(): 168 for rounding in roundings: 169 mod = get_mod(data_shape=(16, ), 170 data_dtype='int32', 171 out_dtype="int8", 172 input_scale=0.5, 173 output_scale=0.5, 174 rounding=rounding) 175 golden_data = np.arange(0, 16, 1).astype('int32') 176 golden_data = np.add(120, golden_data) 177 output = np.array([120, 121, 122, 123, 124, 125, 126, 127, 178 127, 127, 127, 127, 127, 127, 127, 127]) 179 golden_output = output 180 verify(mod, (golden_data, golden_output)) 181 182 # Try negative numbers 183 golden_data = np.arange(0, -16, -1).astype('int32') 184 golden_data = np.add(-120, golden_data) 185 output = np.array([-120, -121, -122, -123, -124, -125, -126, -127, 186 -128, -128, -128, -128, -128, -128, -128, -128]) 187 golden_output = output 188 verify(mod, (golden_data, golden_output)) 189 190def test_zero_point(): 191 # Output zero point 192 for rounding in roundings: 193 mod = get_mod(data_shape=(32, ), 194 data_dtype='int32', 195 out_dtype='int8', 196 input_scale=1, 197 output_scale=16, 198 output_zero_point=1, 199 rounding=rounding) 200 201 # Try positive values 202 # 8 corresponds to 0.5, resulting in 1 203 golden_data = np.arange(0, 32, 1).astype('int32') 204 golden_output = np.repeat([0, 1, 2], [8, 16, 8]) 205 golden_output = np.add(1, golden_output) 206 verify(mod, (golden_data, golden_output)) 207 208 # Try negative values 209 # -8 corresponds to -0.5. For UPWARD, this is 0 210 golden_data = np.arange(-32, -64, -1).astype('int32') 211 if rounding == "UPWARD": 212 golden_output = np.repeat([-2, -3, -4], [9, 16, 7]) 213 else: 214 golden_output = np.repeat([-2, -3, -4], [8, 16, 8]) 215 golden_output = np.add(1, golden_output) 216 verify(mod, (golden_data, golden_output)) 217 218 # Input zero point 219 for rounding in roundings: 220 mod = get_mod(data_shape=(32, ), 221 data_dtype='int32', 222 out_dtype='int8', 223 input_scale=1, 224 output_scale=16, 225 input_zero_point=16, 226 rounding=rounding) 227 228 # Try positive values 229 golden_data = np.arange(32, 64, 1).astype('int32') 230 golden_output = np.repeat([2, 3, 4], [8, 16, 8]) 231 golden_output = np.subtract(golden_output, 1) 232 verify(mod, (golden_data, golden_output)) 233 234 # Try negative values 235 golden_data = np.arange(-32, -64, -1).astype('int32') 236 if rounding == "UPWARD": 237 golden_output = np.repeat([-2, -3, -4], [9, 16, 7]) 238 else: 239 golden_output = np.repeat([-2, -3, -4], [8, 16, 8]) 240 golden_output = np.subtract(golden_output, 1) 241 verify(mod, (golden_data, golden_output)) 242 243if __name__ == "__main__": 244 test_same_scale() 245 test_downscale() 246 test_upscale() 247 test_saturation() 248 test_zero_point() 249