Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 1 | #!/usr/bin/env python |
| 2 | # Copyright 2020 Google LLC |
| 3 | # |
| 4 | # This source code is licensed under the BSD-style license found in the |
| 5 | # LICENSE file in the root directory of this source tree. |
| 6 | |
| 7 | import argparse |
| 8 | import codecs |
| 9 | import os |
| 10 | import re |
| 11 | import sys |
| 12 | import yaml |
| 13 | |
| 14 | sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) |
| 15 | from primes import next_prime |
| 16 | import xngen |
| 17 | import xnncommon |
| 18 | |
| 19 | |
| 20 | parser = argparse.ArgumentParser( |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 21 | description='Test generator for DWCONV2D CHW micro-kernels') |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 22 | parser.add_argument("-s", "--spec", metavar="FILE", required=True, |
| 23 | help="Spec (YAML) file") |
| 24 | parser.add_argument("-o", "--output", metavar="FILE", required=True, |
| 25 | help='Output (C++ source) file') |
| 26 | parser.set_defaults(defines=list()) |
| 27 | |
| 28 | |
| 29 | TEST_TEMPLATE = """\ |
| 30 | $if SUBSAMPLING == 1: |
| 31 | TEST(${TEST_NAME}, output_width_eq_${WIDTH_TILE}) { |
| 32 | $if ISA_CHECK: |
| 33 | ${ISA_CHECK}; |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 34 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 35 | .input_width(${(WIDTH_TILE - 1) * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}) |
| 36 | .input_height(${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}) |
| 37 | .kernel_height(${KERNEL_HEIGHT}) |
| 38 | .kernel_width(${KERNEL_WIDTH}) |
| 39 | .subsampling(${SUBSAMPLING}) |
| 40 | .padding_left(${PADDING}) |
| 41 | .padding_right(${PADDING}) |
| 42 | .padding_top(${PADDING}) |
| 43 | .padding_bottom(${PADDING}) |
| 44 | .Test(${", ".join(TEST_ARGS)}); |
| 45 | } |
| 46 | $else: |
| 47 | TEST(${TEST_NAME}, output_width_eq_${WIDTH_TILE}) { |
| 48 | $if ISA_CHECK: |
| 49 | ${ISA_CHECK}; |
| 50 | for (size_t input_width = ${(WIDTH_TILE - 1) * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width < ${WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width++) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 51 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 52 | .input_width(input_width) |
| 53 | .input_height(${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}) |
| 54 | .kernel_height(${KERNEL_HEIGHT}) |
| 55 | .kernel_width(${KERNEL_WIDTH}) |
| 56 | .subsampling(${SUBSAMPLING}) |
| 57 | .padding_left(${PADDING}) |
| 58 | .padding_right(${PADDING}) |
| 59 | .padding_top(${PADDING}) |
| 60 | .padding_bottom(${PADDING}) |
| 61 | .Test(${", ".join(TEST_ARGS)}); |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | $if WIDTH_TILE > 1: |
| 66 | TEST(${TEST_NAME}, output_width_div_${WIDTH_TILE}) { |
| 67 | $if ISA_CHECK: |
| 68 | ${ISA_CHECK}; |
| 69 | for (size_t input_width = ${2 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING - 1}; input_width < ${8 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING - 1}; input_width += ${WIDTH_TILE * SUBSAMPLING}) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 70 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 71 | .input_width(input_width) |
| 72 | .input_height(${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}) |
| 73 | .kernel_height(${KERNEL_HEIGHT}) |
| 74 | .kernel_width(${KERNEL_WIDTH}) |
| 75 | .subsampling(${SUBSAMPLING}) |
| 76 | .padding_left(${PADDING}) |
| 77 | .padding_right(${PADDING}) |
| 78 | .padding_top(${PADDING}) |
| 79 | .padding_bottom(${PADDING}) |
| 80 | .Test(${", ".join(TEST_ARGS)}); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | TEST(${TEST_NAME}, output_width_lt_${WIDTH_TILE}) { |
| 85 | $if ISA_CHECK: |
| 86 | ${ISA_CHECK}; |
| 87 | for (size_t input_width = ${max(1, KERNEL_WIDTH - 2 * PADDING)}; input_width < ${(WIDTH_TILE - 1) * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width++) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 88 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 89 | .input_width(${WIDTH_TILE * SUBSAMPLING}) |
| 90 | .input_height(${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}) |
| 91 | .kernel_height(${KERNEL_HEIGHT}) |
| 92 | .kernel_width(${KERNEL_WIDTH}) |
| 93 | .subsampling(${SUBSAMPLING}) |
| 94 | .padding_left(${PADDING}) |
| 95 | .padding_right(${PADDING}) |
| 96 | .padding_top(${PADDING}) |
| 97 | .padding_bottom(${PADDING}) |
| 98 | .Test(${", ".join(TEST_ARGS)}); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | TEST(${TEST_NAME}, output_width_gt_${WIDTH_TILE}) { |
| 103 | $if ISA_CHECK: |
| 104 | ${ISA_CHECK}; |
| 105 | for (size_t input_width = ${WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width < ${(5 if WIDTH_TILE == 1 else 2) * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width++) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 106 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 107 | .input_width(input_width) |
| 108 | .input_height(${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}) |
| 109 | .kernel_height(${KERNEL_HEIGHT}) |
| 110 | .kernel_width(${KERNEL_WIDTH}) |
| 111 | .subsampling(${SUBSAMPLING}) |
| 112 | .padding_left(${PADDING}) |
| 113 | .padding_right(${PADDING}) |
| 114 | .padding_top(${PADDING}) |
| 115 | .padding_bottom(${PADDING}) |
| 116 | .Test(${", ".join(TEST_ARGS)}); |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | $if SUBSAMPLING > 1: |
| 121 | TEST(${TEST_NAME}, output_height_eq_${HEIGHT_TILE}) { |
| 122 | $if ISA_CHECK: |
| 123 | ${ISA_CHECK}; |
| 124 | for (size_t input_height = ${(HEIGHT_TILE - 1) * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING}; input_height < ${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING}; input_height++) { |
| 125 | for (size_t input_width = 1; input_width < ${5 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width += ${max(1, WIDTH_TILE * SUBSAMPLING - 1)}) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 126 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 127 | .input_width(input_width) |
| 128 | .input_height(input_height) |
| 129 | .kernel_height(${KERNEL_HEIGHT}) |
| 130 | .kernel_width(${KERNEL_WIDTH}) |
| 131 | .subsampling(${SUBSAMPLING}) |
| 132 | .padding_left(${PADDING}) |
| 133 | .padding_right(${PADDING}) |
| 134 | .padding_top(${PADDING}) |
| 135 | .padding_bottom(${PADDING}) |
| 136 | .Test(${", ".join(TEST_ARGS)}); |
| 137 | } |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | $if HEIGHT_TILE > 1: |
| 142 | TEST(${TEST_NAME}, output_height_div_${HEIGHT_TILE}) { |
| 143 | $if ISA_CHECK: |
| 144 | ${ISA_CHECK}; |
| 145 | for (size_t input_height = ${2 * HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}; input_height < ${8 * HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING - 1}; input_height += ${HEIGHT_TILE * SUBSAMPLING}) { |
| 146 | for (size_t input_width = 1; input_width < ${5 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width += ${max(1, WIDTH_TILE * SUBSAMPLING - 1)}) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 147 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 148 | .input_width(input_width) |
| 149 | .input_height(input_height) |
| 150 | .kernel_height(${KERNEL_HEIGHT}) |
| 151 | .kernel_width(${KERNEL_WIDTH}) |
| 152 | .subsampling(${SUBSAMPLING}) |
| 153 | .padding_left(${PADDING}) |
| 154 | .padding_right(${PADDING}) |
| 155 | .padding_top(${PADDING}) |
| 156 | .padding_bottom(${PADDING}) |
| 157 | .Test(${", ".join(TEST_ARGS)}); |
| 158 | } |
| 159 | } |
| 160 | } |
| 161 | |
| 162 | TEST(${TEST_NAME}, output_height_lt_${HEIGHT_TILE}) { |
| 163 | $if ISA_CHECK: |
| 164 | ${ISA_CHECK}; |
| 165 | for (size_t input_height = ${max(1, KERNEL_HEIGHT - 2 * PADDING)}; input_height < ${(HEIGHT_TILE - 1) * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING}; input_height++) { |
| 166 | for (size_t input_width = 1; input_width < ${5 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width += ${max(1, WIDTH_TILE * SUBSAMPLING - 1)}) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 167 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 168 | .input_width(input_width) |
| 169 | .input_height(input_height) |
| 170 | .kernel_height(${KERNEL_HEIGHT}) |
| 171 | .kernel_width(${KERNEL_WIDTH}) |
| 172 | .subsampling(${SUBSAMPLING}) |
| 173 | .padding_left(${PADDING}) |
| 174 | .padding_right(${PADDING}) |
| 175 | .padding_top(${PADDING}) |
| 176 | .padding_bottom(${PADDING}) |
| 177 | .Test(${", ".join(TEST_ARGS)}); |
| 178 | } |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | TEST(${TEST_NAME}, output_height_gt_${HEIGHT_TILE}) { |
| 183 | $if ISA_CHECK: |
| 184 | ${ISA_CHECK}; |
| 185 | for (size_t input_height = ${HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING}; input_height < ${(5 if WIDTH_TILE == 1 else 2) * HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING}; input_height++) { |
| 186 | for (size_t input_width = 1; input_width < ${5 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width += ${max(1, WIDTH_TILE * SUBSAMPLING - 1)}) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 187 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 188 | .input_width(input_width) |
| 189 | .input_height(input_height) |
| 190 | .kernel_height(${KERNEL_HEIGHT}) |
| 191 | .kernel_width(${KERNEL_WIDTH}) |
| 192 | .subsampling(${SUBSAMPLING}) |
| 193 | .padding_left(${PADDING}) |
| 194 | .padding_right(${PADDING}) |
| 195 | .padding_top(${PADDING}) |
| 196 | .padding_bottom(${PADDING}) |
| 197 | .Test(${", ".join(TEST_ARGS)}); |
| 198 | } |
| 199 | } |
| 200 | } |
| 201 | |
| 202 | $if SUBSAMPLING > 1: |
| 203 | TEST(${TEST_NAME}, padding_top_eq_${SUBSAMPLING - 1}) { |
| 204 | $if ISA_CHECK: |
| 205 | ${ISA_CHECK}; |
| 206 | for (size_t input_height = ${max(1, KERNEL_HEIGHT - 2 * PADDING + 1)}; input_height < ${3 * HEIGHT_TILE * SUBSAMPLING + KERNEL_HEIGHT - 2 * PADDING + 1}; input_height++) { |
| 207 | for (size_t input_width = 1; input_width < ${5 * WIDTH_TILE * SUBSAMPLING + KERNEL_WIDTH - 2 * PADDING}; input_width += ${max(1, WIDTH_TILE * SUBSAMPLING - 1)}) { |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 208 | DWConv2DMicrokernelTester() |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 209 | .input_width(input_width) |
| 210 | .input_height(input_height) |
| 211 | .kernel_height(${KERNEL_HEIGHT}) |
| 212 | .kernel_width(${KERNEL_WIDTH}) |
| 213 | .subsampling(${SUBSAMPLING}) |
| 214 | .padding_left(${PADDING}) |
| 215 | .padding_right(${PADDING}) |
| 216 | .padding_top(${PADDING - 1}) |
| 217 | .padding_bottom(${PADDING}) |
| 218 | .Test(${", ".join(TEST_ARGS)}); |
| 219 | } |
| 220 | } |
| 221 | } |
| 222 | """ |
| 223 | |
| 224 | def split_ukernel_name(name): |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 225 | match = re.match(r"^xnn_(f16|f32)_dwconv2d_chw_ukernel_(\d+)x(\d+)(s2)?p(\d+)__(.+)_(\d+)x(\d+)(_acc\d+)?$", name) |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 226 | assert match is not None |
| 227 | kernel_height, kernel_width = int(match.group(2)), int(match.group(3)) |
| 228 | if match.group(4): |
| 229 | assert match.group(4).startswith("s") |
| 230 | stride = int(match.group(4)[1:]) |
| 231 | else: |
| 232 | stride = 1 |
| 233 | padding = int(match.group(5)) |
| 234 | |
| 235 | height_tile = int(match.group(7)) |
| 236 | width_tile = int(match.group(8)) |
| 237 | |
| 238 | arch, isa = xnncommon.parse_target_name(target_name=match.group(6)) |
| 239 | return kernel_height, kernel_width, stride, padding, arch, isa, \ |
| 240 | height_tile, width_tile |
| 241 | |
| 242 | |
| 243 | def generate_test_cases(ukernel, kernel_height, kernel_width, subsampling, \ |
| 244 | padding, isa, height_tile, width_tile): |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 245 | """Generates all tests cases for a DWCONV2D CHW micro-kernel. |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 246 | |
| 247 | Args: |
| 248 | ukernel: C name of the micro-kernel function. |
| 249 | kernel_height: convolution kernel height assumed by the micro-kernel. |
| 250 | kernel_width: convolution kernel width assumed by the micro-kernel. |
| 251 | subsampling: convolution subsampling (stride) assumed by the micro-kernel. |
| 252 | The same subsampling factor is assumed for both horizontal and |
| 253 | vertical directions. |
| 254 | padding: convolution padding value assumed by the micro-kernel for right, |
| 255 | bottom, and left padding. If convolution stride is 1, the same |
| 256 | padding value is assumed for the top padding. If convolution stride |
| 257 | is different than 1, top padding is specified via micro-kernel |
| 258 | parameter, and can be either padding or (padding - 1). |
| 259 | isa: instruction set required to run the micro-kernel. Generated unit test |
| 260 | will skip execution if the host processor doesn't support this ISA. |
| 261 | height_tile: number of output rows processed in one iteration of the main |
| 262 | loop of the micro-kernel. |
| 263 | width_tile: number of output columns processed in one iteration of the main |
| 264 | loop of the micro-kernel. |
| 265 | |
| 266 | Returns: |
| 267 | Code for the test case. |
| 268 | """ |
| 269 | _, test_name = ukernel.split("_", 1) |
| 270 | _, datatype, ukernel_type, _ = ukernel.split("_", 3) |
| 271 | test_args = [ukernel] |
| 272 | if not isa or isa == "psimd": |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 273 | test_args.append("DWConv2DMicrokernelTester::Variant::Scalar") |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 274 | return xngen.preprocess(TEST_TEMPLATE, { |
| 275 | "TEST_NAME": test_name.upper().replace("UKERNEL_", ""), |
| 276 | "TEST_ARGS": test_args, |
| 277 | "UKERNEL_TYPE": ukernel_type.upper(), |
| 278 | "DATATYPE": datatype, |
| 279 | "KERNEL_HEIGHT": kernel_height, |
| 280 | "KERNEL_WIDTH": kernel_width, |
| 281 | "SUBSAMPLING": subsampling, |
| 282 | "PADDING": padding, |
| 283 | "HEIGHT_TILE": height_tile, |
| 284 | "WIDTH_TILE": width_tile, |
| 285 | "ISA_CHECK": xnncommon.generate_isa_check_macro(isa), |
| 286 | "next_prime": next_prime, |
| 287 | }) |
| 288 | |
| 289 | |
| 290 | def main(args): |
| 291 | options = parser.parse_args(args) |
| 292 | |
| 293 | with codecs.open(options.spec, "r", encoding="utf-8") as spec_file: |
| 294 | spec_yaml = yaml.safe_load(spec_file) |
| 295 | if not isinstance(spec_yaml, list): |
| 296 | raise ValueError("expected a list of micro-kernels in the spec") |
| 297 | |
| 298 | tests = """\ |
| 299 | // Copyright 2020 Google LLC |
| 300 | // |
| 301 | // This source code is licensed under the BSD-style license found in the |
| 302 | // LICENSE file in the root directory of this source tree. |
| 303 | // |
| 304 | // Auto-generated file. Do not edit! |
| 305 | // Specification: {specification} |
| 306 | // Generator: {generator} |
| 307 | |
| 308 | |
| 309 | #include <gtest/gtest.h> |
| 310 | |
| 311 | #include <xnnpack/common.h> |
| 312 | #include <xnnpack/isa-checks.h> |
| 313 | |
| 314 | #include <xnnpack/dwconv.h> |
Marat Dukhan | bf715f9 | 2020-10-23 20:17:00 -0700 | [diff] [blame] | 315 | #include "dwconv2d-microkernel-tester.h" |
Marat Dukhan | dc6c77f | 2020-10-23 19:09:10 -0700 | [diff] [blame] | 316 | """.format(specification=options.spec, generator=sys.argv[0]) |
| 317 | |
| 318 | for ukernel_spec in spec_yaml: |
| 319 | name = ukernel_spec["name"] |
| 320 | pipelined = bool(ukernel_spec.get("pipelined", False)) |
| 321 | assembly = bool(ukernel_spec.get("assembly", False)) |
| 322 | kernel_height, kernel_width, subsampling, padding, arch, isa, \ |
| 323 | height_tile, width_tile = split_ukernel_name(name) |
| 324 | |
| 325 | # specification can override architecture |
| 326 | arch = ukernel_spec.get("arch", arch) |
| 327 | |
| 328 | test_case = generate_test_cases(name, kernel_height, kernel_width, \ |
| 329 | subsampling, padding, isa, \ |
| 330 | height_tile, width_tile) |
| 331 | tests += "\n\n" + xnncommon.postprocess_test_case(test_case, arch, isa, assembly) |
| 332 | |
| 333 | with codecs.open(options.output, "w", encoding="utf-8") as output_file: |
| 334 | output_file.write(tests) |
| 335 | |
| 336 | |
| 337 | if __name__ == "__main__": |
| 338 | main(sys.argv[1:]) |