|
| 1 | +# Copyright (c) Qualcomm Innovation Center, Inc. |
| 2 | +# All rights reserved. |
| 3 | +# |
| 4 | +# SPDX-License-Identifier: BSD-3-Clause |
| 5 | + |
| 6 | +""" |
| 7 | +Example: Custom FastGELU operator integrated with ExecuTorch Qualcomm backend (HTP). |
| 8 | +""" |
| 9 | + |
| 10 | +import json |
| 11 | +import os |
| 12 | +import subprocess |
| 13 | +import sys |
| 14 | +from multiprocessing.connection import Client |
| 15 | + |
| 16 | +import numpy as np |
| 17 | +import torch |
| 18 | + |
| 19 | +from executorch.backends.qualcomm.quantizer.quantizer import QuantDtype |
| 20 | +from executorch.backends.qualcomm.serialization.qc_schema import ( |
| 21 | + _soc_info_table, |
| 22 | + HtpArch, |
| 23 | + QcomChipset, |
| 24 | + QnnExecuTorchOpPackageInfo, |
| 25 | + QnnExecuTorchOpPackageOptions, |
| 26 | + QnnExecuTorchOpPackagePlatform, |
| 27 | + QnnExecuTorchOpPackageTarget, |
| 28 | +) |
| 29 | +from executorch.examples.qualcomm.utils import ( |
| 30 | + build_executorch_binary, |
| 31 | + generate_inputs, |
| 32 | + make_output_dir, |
| 33 | + make_quantizer, |
| 34 | + setup_common_args_and_variables, |
| 35 | + SimpleADB, |
| 36 | +) |
| 37 | +from torch.library import impl, Library |
| 38 | + |
| 39 | +# ------------------------------------------------------------------------------ |
| 40 | +# 1. Register PyTorch custom operator (FastGELU) |
| 41 | +# ------------------------------------------------------------------------------ |
| 42 | + |
| 43 | +my_op_lib = Library("my_ops", "DEF") |
| 44 | +my_op_lib.define("fast_gelu(Tensor input) -> Tensor") |
| 45 | + |
| 46 | + |
| 47 | +@impl(my_op_lib, "fast_gelu", "CompositeExplicitAutograd") |
| 48 | +def fast_gelu_impl(x: torch.Tensor) -> torch.Tensor: |
| 49 | + return 0.5 * x * (1.0 + torch.tanh(0.7978845608 * (x + 0.044715 * x * x * x))) |
| 50 | + |
| 51 | + |
| 52 | +# registering the out variant. |
| 53 | +my_op_lib.define( |
| 54 | + "fast_gelu.out(Tensor input, *, Tensor(a!) output) -> Tensor(a!)" |
| 55 | +) # should print 'fast_gelu.out' |
| 56 | + |
| 57 | + |
| 58 | +# ------------------------------------------------------------------------------ |
| 59 | +# 2. Simple model using custom op |
| 60 | +# ------------------------------------------------------------------------------ |
| 61 | + |
| 62 | + |
| 63 | +class Model(torch.nn.Module): |
| 64 | + def forward(self, a): |
| 65 | + return torch.ops.my_ops.fast_gelu.default(a) |
| 66 | + |
| 67 | + |
| 68 | +# ------------------------------------------------------------------------------ |
| 69 | +# 3. Build + register custom op package |
| 70 | +# ------------------------------------------------------------------------------ |
| 71 | + |
| 72 | + |
| 73 | +def _run(cmd, cwd=None): |
| 74 | + subprocess.run(cmd, stdout=sys.stdout, cwd=cwd, check=True) |
| 75 | + |
| 76 | + |
| 77 | +def prepare_op_package( |
| 78 | + workspace: str, op_package_dir: str, arch: HtpArch, build_op_package: bool |
| 79 | +): |
| 80 | + if build_op_package: |
| 81 | + _run(["rm", "-rf", "build"], cwd=op_package_dir) |
| 82 | + _run(["make", "htp_x86", "htp_aarch64", f"htp_v{arch}"], cwd=op_package_dir) |
| 83 | + _run( |
| 84 | + [ |
| 85 | + "cp", |
| 86 | + f"{op_package_dir}/build/hexagon-v{arch}/libQnnFastGeluOpPackage.so", |
| 87 | + f"{op_package_dir}/build/hexagon-v{arch}/libQnnFastGeluOpPackage_HTP.so", |
| 88 | + ] |
| 89 | + ) |
| 90 | + |
| 91 | + op_package_paths = [ |
| 92 | + f"{op_package_dir}/build/hexagon-v{arch}/libQnnFastGeluOpPackage_HTP.so", |
| 93 | + f"{op_package_dir}/build/aarch64-android/libQnnFastGeluOpPackage.so", |
| 94 | + ] |
| 95 | + |
| 96 | + op_package_infos_HTP = QnnExecuTorchOpPackageInfo() |
| 97 | + op_package_infos_HTP.interface_provider = "FastGeluOpPackageInterfaceProvider" |
| 98 | + op_package_infos_HTP.op_package_name = "FastGeluOpPackage" |
| 99 | + op_package_infos_HTP.op_package_path = f"{workspace}/libQnnFastGeluOpPackage_HTP.so" |
| 100 | + op_package_infos_HTP.target = QnnExecuTorchOpPackageTarget.HTP |
| 101 | + op_package_infos_HTP.custom_op_name = "my_ops.fast_gelu.default" |
| 102 | + op_package_infos_HTP.qnn_op_type_name = "FastGelu" |
| 103 | + op_package_infos_HTP.platform = QnnExecuTorchOpPackagePlatform.AARCH64_ANDROID |
| 104 | + op_package_infos_aarch64_CPU = QnnExecuTorchOpPackageInfo() |
| 105 | + op_package_infos_aarch64_CPU.interface_provider = ( |
| 106 | + "FastGeluOpPackageInterfaceProvider" |
| 107 | + ) |
| 108 | + op_package_infos_aarch64_CPU.op_package_name = "FastGeluOpPackage" |
| 109 | + op_package_infos_aarch64_CPU.op_package_path = ( |
| 110 | + f"{workspace}/libQnnFastGeluOpPackage.so" |
| 111 | + ) |
| 112 | + op_package_infos_aarch64_CPU.target = QnnExecuTorchOpPackageTarget.CPU |
| 113 | + op_package_infos_aarch64_CPU.custom_op_name = "my_ops.fast_gelu.default" |
| 114 | + op_package_infos_aarch64_CPU.qnn_op_type_name = "FastGelu" |
| 115 | + op_package_infos_aarch64_CPU.platform = ( |
| 116 | + QnnExecuTorchOpPackagePlatform.AARCH64_ANDROID |
| 117 | + ) |
| 118 | + op_package_infos_x86_CPU = QnnExecuTorchOpPackageInfo() |
| 119 | + op_package_infos_x86_CPU.interface_provider = "FastGeluOpPackageInterfaceProvider" |
| 120 | + op_package_infos_x86_CPU.op_package_name = "FastGeluOpPackage" |
| 121 | + op_package_infos_x86_CPU.op_package_path = ( |
| 122 | + f"{op_package_dir}/build/x86_64-linux-clang/libQnnFastGeluOpPackage.so" |
| 123 | + ) |
| 124 | + op_package_infos_x86_CPU.target = QnnExecuTorchOpPackageTarget.CPU |
| 125 | + op_package_infos_x86_CPU.custom_op_name = "my_ops.fast_gelu.default" |
| 126 | + op_package_infos_x86_CPU.qnn_op_type_name = "FastGelu" |
| 127 | + op_package_infos_x86_CPU.platform = QnnExecuTorchOpPackagePlatform.X86_64 |
| 128 | + op_package_options = QnnExecuTorchOpPackageOptions() |
| 129 | + op_package_options.op_package_infos = [ |
| 130 | + op_package_infos_x86_CPU, |
| 131 | + op_package_infos_aarch64_CPU, |
| 132 | + op_package_infos_HTP, |
| 133 | + ] |
| 134 | + |
| 135 | + return op_package_options, op_package_paths |
| 136 | + |
| 137 | + |
| 138 | +# ------------------------------------------------------------------------------ |
| 139 | +# 4. Entrypoint — same pattern as custom_ops_1.py |
| 140 | +# ------------------------------------------------------------------------------ |
| 141 | + |
| 142 | + |
| 143 | +def main(args): |
| 144 | + if args.build_op_package: |
| 145 | + if "HEXAGON_SDK_ROOT" not in os.environ: |
| 146 | + raise RuntimeError("Environment variable HEXAGON_SDK_ROOT must be set") |
| 147 | + print(f"HEXAGON_SDK_ROOT={os.getenv('HEXAGON_SDK_ROOT')}") |
| 148 | + |
| 149 | + if "ANDROID_NDK_ROOT" not in os.environ: |
| 150 | + raise RuntimeError("Environment variable ANDROID_NDK_ROOT must be set") |
| 151 | + print(f"ANDROID_NDK_ROOT={os.getenv('ANDROID_NDK_ROOT')}") |
| 152 | + |
| 153 | + # ensure the working directory exist. |
| 154 | + os.makedirs(args.artifact, exist_ok=True) |
| 155 | + |
| 156 | + quant_dtype = QuantDtype.use_8a8w |
| 157 | + if args.use_fp16: |
| 158 | + quant_dtype = None |
| 159 | + |
| 160 | + instance = Model() |
| 161 | + sample_input = (torch.randn(1, 128),) |
| 162 | + pte_filename = "fastgelu_model" |
| 163 | + workspace = f"/data/local/tmp/executorch/{pte_filename}" |
| 164 | + soc_info: SocInfo = _soc_info_table[getattr(QcomChipset, args.model)] |
| 165 | + |
| 166 | + op_package_options, op_package_paths = prepare_op_package( |
| 167 | + workspace, |
| 168 | + args.op_package_dir, |
| 169 | + soc_info.htp_info.htp_arch, |
| 170 | + args.build_op_package, |
| 171 | + ) |
| 172 | + # quantizer = make_quantizer( |
| 173 | + # quant_dtype=quant_dtype, custom_annotations=(annotate_custom,) |
| 174 | + # ) |
| 175 | + |
| 176 | + build_executorch_binary( |
| 177 | + instance, |
| 178 | + sample_input, |
| 179 | + args.model, |
| 180 | + f"{args.artifact}/{pte_filename}", |
| 181 | + sample_input, |
| 182 | + op_package_options=op_package_options, |
| 183 | + # quant_dtype=quant_dtype, |
| 184 | + # custom_quantizer=quantizer, |
| 185 | + ) |
| 186 | + |
| 187 | + if args.compile_only: |
| 188 | + sys.exit(0) |
| 189 | + |
| 190 | + output_data_folder: LiteralString = os.path.join(args.artifact, "outputs") |
| 191 | + |
| 192 | + adb = SimpleADB( |
| 193 | + qnn_sdk=os.getenv("QNN_SDK_ROOT"), |
| 194 | + build_path=args.build_folder, |
| 195 | + pte_path=f"{args.artifact}/{pte_filename}.pte", |
| 196 | + workspace=workspace, |
| 197 | + device_id=args.device, |
| 198 | + host_id=args.host, |
| 199 | + soc_model=args.model, |
| 200 | + ) |
| 201 | + adb.push(inputs=sample_input, files=op_package_paths) |
| 202 | + adb.execute() |
| 203 | + adb.pull(output_path=args.artifact) |
| 204 | + |
| 205 | + # Compare results |
| 206 | + x86_golden = model(*sample_input) |
| 207 | + import numpy as np |
| 208 | + |
| 209 | + device_output = torch.from_numpy( |
| 210 | + np.fromfile( |
| 211 | + os.path.join(output_data_folder, "output_0_0.raw"), dtype=np.float32 |
| 212 | + ) |
| 213 | + ).reshape(x86_golden.size()) |
| 214 | + print( |
| 215 | + "is_close?", |
| 216 | + torch.all(torch.isclose(x86_golden, device_output, atol=1e-2)).item(), |
| 217 | + ) |
| 218 | + |
| 219 | + |
| 220 | +if __name__ == "__main__": |
| 221 | + parser = setup_common_args_and_variables() |
| 222 | + parser.add_argument( |
| 223 | + "-a", |
| 224 | + "--artifact", |
| 225 | + help="path for storing generated artifacts by this example. Default ./custom_op", |
| 226 | + default="./custom_op", |
| 227 | + type=str, |
| 228 | + ) |
| 229 | + |
| 230 | + parser.add_argument( |
| 231 | + "-d", |
| 232 | + "--op_package_dir", |
| 233 | + help="Path to operator package which generates from QNN.", |
| 234 | + type=str, |
| 235 | + required=True, |
| 236 | + ) |
| 237 | + |
| 238 | + parser.add_argument( |
| 239 | + "-F", |
| 240 | + "--use_fp16", |
| 241 | + help="If specified, will run in fp16 precision and discard ptq setting", |
| 242 | + action="store_true", |
| 243 | + default=False, |
| 244 | + ) |
| 245 | + |
| 246 | + parser.add_argument( |
| 247 | + "--build_op_package", |
| 248 | + help="Build op package based on op_package_dir. Please set up " |
| 249 | + "`HEXAGON_SDK_ROOT` and `ANDROID_NDK_ROOT` environment variable. " |
| 250 | + "And add clang compiler into `PATH`. Please refer to Qualcomm AI Engine " |
| 251 | + "Direct SDK document to get more details", |
| 252 | + action="store_true", |
| 253 | + default=False, |
| 254 | + ) |
| 255 | + |
| 256 | + args = parser.parse_args() |
| 257 | + args.validate(args) |
| 258 | + |
| 259 | + try: |
| 260 | + main(args) |
| 261 | + except Exception as e: |
| 262 | + if args.ip and args.port != -1: |
| 263 | + with Client((args.ip, args.port)) as conn: |
| 264 | + conn.send(json.dumps({"Error": str(e)})) |
| 265 | + else: |
| 266 | + raise Exception(e) |
0 commit comments