iswaalex commited on
Commit
b39425e
·
verified ·
1 Parent(s): 77c7fd4

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. benchmark_threading.py +302 -0
  2. segmentation_1_4_0_fp32_combined/aie_unsupported_original_ops.json +13 -0
  3. segmentation_1_4_0_fp32_combined/cache/1d4c9c71f3a11a5a3ebd0c4f9b9709fb907ad630/te_inter_partition_cache.json +0 -0
  4. segmentation_1_4_0_fp32_combined/context.json +1052 -0
  5. segmentation_1_4_0_fp32_combined/final-vaiml-pass-summary.txt +23 -0
  6. segmentation_1_4_0_fp32_combined/gops.csv +318 -0
  7. segmentation_1_4_0_fp32_combined/graph_partition_trace.csv +318 -0
  8. segmentation_1_4_0_fp32_combined/original-info-signature.txt +1 -0
  9. segmentation_1_4_0_fp32_combined/original-model-signature.txt +1 -0
  10. segmentation_1_4_0_fp32_combined/preliminary-vaiml-pass-summary.txt +14 -0
  11. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/AIECompiler.log +0 -0
  12. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/.target +1 -0
  13. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/0_0.log +59 -0
  14. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0 +0 -0
  15. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.# +2 -0
  16. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.## +2 -0
  17. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.calltree +32 -0
  18. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.cmic2 +0 -0
  19. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.cmico +1 -0
  20. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.lst +824 -0
  21. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.map +143 -0
  22. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.o +0 -0
  23. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.o.lst +471 -0
  24. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.sdr +90 -0
  25. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/elf_ctrl_pkt.bin +3 -0
  26. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_0.bin +3 -0
  27. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_1.bin +3 -0
  28. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_100.bin +3 -0
  29. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_11.bin +3 -0
  30. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/xlopt.log +81 -0
  31. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/AddressSpace.txt +0 -0
  32. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/AliasAnalysisReport.txt +0 -0
  33. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/Makefile +0 -0
  34. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/layer_control_parameters.json +0 -0
  35. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/top.aiecompile_summary +0 -0
  36. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/aie.mk +2 -0
  37. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/aiecompiler-flexml.log +0 -0
  38. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/aiesim_cmd_opts.txt +0 -0
  39. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/xcd.log +30 -0
  40. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/buffer_info.json +0 -0
  41. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/ctrlPktPatch.json +0 -0
  42. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/external_buffer_id.json +0 -0
  43. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/flexmlrt-hsi.json +1253 -0
  44. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/ml_txn.bin +3 -0
  45. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/model_top.cpp +6 -0
  46. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/par.mlopslib.tosa.mlir +0 -0
  47. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/par.subgraph.dse.mlir +0 -0
  48. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/par.subgraph.pre-dse.mlir +0 -0
  49. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/tensor_dims.json +0 -0
  50. segmentation_1_4_0_fp32_combined/vaiml_par_0/0/unified-4x4.xclbin +0 -0
benchmark_threading.py ADDED
@@ -0,0 +1,302 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import time
3
+ import argparse
4
+ import subprocess
5
+ import platform
6
+ from typing import Optional, Tuple, Dict
7
+ import threading
8
+ import numpy as np
9
+ from onnx import load, ModelProto
10
+ import onnxruntime as ort
11
+
12
+ os.environ["XLNX_ENABLE_CACHE"] = "0"
13
+ os.environ["PATH"] += (
14
+ os.pathsep + f"{os.environ['CONDA_PREFIX']}\\Lib\\site-packages\\flexmlrt\\lib"
15
+ )
16
+
17
+ XRT_SMI_PATH = "C:\\Windows\\System32\\AMD\\xrt-smi.exe"
18
+
19
+ ONNX_DTYPE_TO_NP = {
20
+ "tensor(float)": np.float32,
21
+ "tensor(float16)": np.float16,
22
+ "tensor(int64)": np.int64,
23
+ "tensor(int32)": np.int32,
24
+ "tensor(uint16)": np.uint16,
25
+ "tensor(int16)": np.int16,
26
+ "tensor(uint8)": np.uint8,
27
+ "tensor(int8)": np.int8,
28
+ }
29
+
30
+
31
+ def generate_rand_data_from_onnx(
32
+ model: ModelProto,
33
+ lowest_int_val: Optional[int],
34
+ highest_int_val: Optional[int],
35
+ ) -> Dict[str, np.ndarray]:
36
+ # Load the models
37
+
38
+ sess = ort.InferenceSession(
39
+ model.SerializePartialToString(), providers=["CPUExecutionProvider"]
40
+ )
41
+
42
+ inps = {}
43
+ # Iterate over the first models inputs and generate random data
44
+ for inp in sess.get_inputs():
45
+ inp_shapes = list(inp.shape) # mutable
46
+ for inp_shape in inp_shapes:
47
+ assert isinstance(
48
+ inp_shape, int
49
+ ), f"Found dynamic axes: {inp_shape}. Please freeze."
50
+ np_type = ONNX_DTYPE_TO_NP[inp.type]
51
+ if np.issubdtype(np_type, np.integer):
52
+ iinfo = np.iinfo(np_type)
53
+ if lowest_int_val is None:
54
+ lowest_int_val = iinfo.min
55
+ if highest_int_val is None:
56
+ lowest_int_val = iinfo.max
57
+ inps[inp.name] = np.random.randint(
58
+ lowest_int_val, highest_int_val, size=tuple(inp_shapes), dtype=np_type
59
+ )
60
+ else:
61
+ inps[inp.name] = np.random.rand(*inp_shapes).astype(np_type)
62
+
63
+ return inps
64
+
65
+
66
+ def configure_npu_power(p_mode: Optional[str] = None) -> Tuple[int, str, str]:
67
+ """
68
+ Configures the NPU power state using xrt-smi.exe.
69
+
70
+ Args:
71
+ p_mode (string, optional): The desired power mode (p-mode).
72
+ If None, displays current status.
73
+ Refer to xrt-smi documentation for valid p-modes.
74
+ Returns:
75
+ tuple: (return_code, stdout, stderr) from the subprocess call.
76
+ return_code is an integer, stdout and stderr are strings.
77
+ Raises:
78
+ OSError: If xrt-smi.exe is not found.
79
+ """
80
+
81
+ if platform.system() != "Windows":
82
+ return (-1, "xrt-smi.exe is only available on Windows.", "")
83
+
84
+ try:
85
+ if p_mode is not None:
86
+ command = [XRT_SMI_PATH, "configure", "--pmode", str(p_mode)]
87
+ else:
88
+ command = [
89
+ XRT_SMI_PATH,
90
+ "examine",
91
+ "--report",
92
+ "platform",
93
+ ] # Just display status
94
+
95
+ process = subprocess.Popen(
96
+ command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True
97
+ )
98
+ stdout, stderr = process.communicate()
99
+ return_code = process.returncode
100
+
101
+ if return_code != 0:
102
+ print(f"Error executing xrt-smi.exe: {stderr}")
103
+
104
+ return return_code, stdout, stderr
105
+
106
+ except FileNotFoundError:
107
+ raise OSError("xrt-smi.exe not found.")
108
+ except Exception as e: # pylint: disable=broad-except
109
+ print(f"An unexpected error occurred: {e}")
110
+ return -1, "", str(e)
111
+
112
+
113
+ def main(
114
+ model_file: str,
115
+ vaip_config: str,
116
+ cache_path: str,
117
+ device: str,
118
+ pmode: str,
119
+ warmup_runs: int,
120
+ inferences: int,
121
+ lowest_int_value: Optional[int],
122
+ highest_int_value: Optional[int],
123
+ threads: int,
124
+ ):
125
+
126
+ assert os.path.exists(model_file)
127
+
128
+ assert threads >= 1
129
+
130
+ if device == "cpu":
131
+ ort_session = ort.InferenceSession(
132
+ model_file,
133
+ providers=["CPUExecutionProvider"],
134
+ )
135
+
136
+ elif device == "npu":
137
+ assert os.path.exists(vaip_config)
138
+ assert os.path.exists(cache_path)
139
+ cache_dir = os.path.dirname(os.path.abspath(cache_path))
140
+ cache_key = os.path.basename(cache_path)
141
+ print(f"Using cache directory {cache_dir} with key {cache_key}")
142
+
143
+ ret_code, stdout, stderr = configure_npu_power(pmode)
144
+ print(stdout)
145
+ if ret_code != 0:
146
+ print("Error configuring npu power mode.")
147
+ print(stderr)
148
+
149
+ sess_options = ort.SessionOptions()
150
+ ort_session = ort.InferenceSession(
151
+ model_file,
152
+ providers=["VitisAIExecutionProvider"],
153
+ sess_options=sess_options,
154
+ provider_options=[
155
+ {
156
+ "config_file": vaip_config,
157
+ "cacheDir": cache_dir,
158
+ "cacheKey": cache_key,
159
+ }
160
+ ],
161
+ )
162
+
163
+ elif device == "igpu":
164
+ ort_session = ort.InferenceSession(
165
+ model_file,
166
+ providers=["DmlExecutionProvider"],
167
+ provider_options=[{"device_id": 2}],
168
+ )
169
+
170
+ onnx_inputs = generate_rand_data_from_onnx(
171
+ load(model_file),
172
+ lowest_int_val=lowest_int_value,
173
+ highest_int_val=highest_int_value,
174
+ )
175
+
176
+ # Warmup
177
+ for _ in range(warmup_runs):
178
+ ort_session.run(None, onnx_inputs)
179
+
180
+ def run_inference(runs, session, inputs):
181
+ for _ in range(runs):
182
+ session.run(None, inputs)
183
+
184
+ latencies = []
185
+ num_threads = threads
186
+ threads_list = []
187
+ inferences_per_thread = inferences // num_threads
188
+ remainder = inferences % num_threads
189
+ print(f"inferences per thread: {inferences_per_thread}, remainder: {remainder}")
190
+ start = time.perf_counter()
191
+ for i in range(num_threads):
192
+ num_runs = inferences_per_thread + (1 if i < remainder else 0)
193
+ thread = threading.Thread(
194
+ target=run_inference, args=(num_runs, ort_session, onnx_inputs)
195
+ )
196
+ threads_list.append(thread)
197
+ thread.start()
198
+
199
+ for thread in threads_list:
200
+ thread.join()
201
+
202
+ end = time.perf_counter()
203
+ latencies.append((end - start) / inferences)
204
+ print(f"Latencies: {latencies}")
205
+ print(f"Benchmark results averaged over {inferences} inferences targeting {device}")
206
+ print("Average latency (ms): ", round(np.mean(latencies) * 1e3, 3))
207
+ print("Average throughput (inf/s): ", round(1 / np.mean(latencies), 3))
208
+
209
+
210
+ if __name__ == "__main__":
211
+ parser = argparse.ArgumentParser(
212
+ description="",
213
+ )
214
+
215
+ parser.add_argument(
216
+ "--pmode",
217
+ type=str,
218
+ choices=["default", "powersaver", "balanced", "performance", "turbo"],
219
+ default="default",
220
+ help="Desired power mode.",
221
+ )
222
+
223
+ parser.add_argument(
224
+ "onnx_model",
225
+ type=str,
226
+ help="Provide the onnx model file.",
227
+ )
228
+
229
+ parser.add_argument(
230
+ "--vai-config",
231
+ type=str,
232
+ help="Path to the vaip configuration json file.",
233
+ )
234
+
235
+ parser.add_argument(
236
+ "--cache-path",
237
+ required=False,
238
+ type=str,
239
+ help="Path to the saved compilation directory.",
240
+ )
241
+
242
+ parser.add_argument(
243
+ "--device",
244
+ required=False,
245
+ type=str,
246
+ default="npu",
247
+ choices=["npu", "cpu", "igpu"],
248
+ help="Select the device to run the measurements on.",
249
+ )
250
+
251
+ parser.add_argument(
252
+ "--warmup-runs",
253
+ required=False,
254
+ default=10,
255
+ type=int,
256
+ help="The number of inferences to run before capturing performance.",
257
+ )
258
+
259
+ parser.add_argument(
260
+ "--inferences",
261
+ required=False,
262
+ default=100,
263
+ type=int,
264
+ help="The number of inferences to average performance over.",
265
+ )
266
+
267
+ parser.add_argument(
268
+ "--lowest-int-value",
269
+ required=False,
270
+ type=int,
271
+ help="Lowest value the rng will produce if the model has an integer input type.",
272
+ )
273
+
274
+ parser.add_argument(
275
+ "--highest-int-value",
276
+ required=False,
277
+ type=int,
278
+ help="Highest value the rng will produce if the model has an integer input type.",
279
+ )
280
+
281
+ parser.add_argument(
282
+ "--threads",
283
+ default=1,
284
+ required=False,
285
+ type=int,
286
+ help="The number of threads that are used to run the inferences.",
287
+ )
288
+
289
+ args = parser.parse_args()
290
+
291
+ main(
292
+ args.onnx_model,
293
+ args.vai_config,
294
+ args.cache_path,
295
+ args.device,
296
+ args.pmode,
297
+ args.warmup_runs,
298
+ args.inferences,
299
+ args.lowest_int_value,
300
+ args.highest_int_value,
301
+ args.threads,
302
+ )
segmentation_1_4_0_fp32_combined/aie_unsupported_original_ops.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ "Cast_0",
3
+ "Transpose_10",
4
+ "Transpose_11",
5
+ "Transpose_12",
6
+ "Transpose_448",
7
+ "Transpose_449",
8
+ "Transpose_450",
9
+ "Transpose_451",
10
+ "Transpose_452",
11
+ "Transpose_453",
12
+ "Transpose_9"
13
+ ]
segmentation_1_4_0_fp32_combined/cache/1d4c9c71f3a11a5a3ebd0c4f9b9709fb907ad630/te_inter_partition_cache.json ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/context.json ADDED
@@ -0,0 +1,1052 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metaDef": [
3
+ {
4
+ "id": "vaiml_par_0",
5
+ "inputs": [
6
+ "385",
7
+ "394",
8
+ "395",
9
+ "396",
10
+ "397"
11
+ ],
12
+ "outputs": [
13
+ "921",
14
+ "894",
15
+ "868",
16
+ "832",
17
+ "796",
18
+ "916"
19
+ ],
20
+ "nodes": [
21
+ "921",
22
+ "909",
23
+ "908",
24
+ "907",
25
+ "1081",
26
+ "904",
27
+ "1078",
28
+ "901",
29
+ "900",
30
+ "895",
31
+ "894",
32
+ "893",
33
+ "889",
34
+ "888",
35
+ "887",
36
+ "886",
37
+ "884",
38
+ "883",
39
+ "882",
40
+ "881",
41
+ "879",
42
+ "878",
43
+ "1075",
44
+ "875",
45
+ "874",
46
+ "869",
47
+ "868",
48
+ "867",
49
+ "863",
50
+ "862",
51
+ "861",
52
+ "860",
53
+ "858",
54
+ "857",
55
+ "856",
56
+ "855",
57
+ "853",
58
+ "852",
59
+ "1072",
60
+ "849",
61
+ "848",
62
+ "838",
63
+ "833",
64
+ "832",
65
+ "831",
66
+ "827",
67
+ "826",
68
+ "825",
69
+ "824",
70
+ "822",
71
+ "821",
72
+ "820",
73
+ "819",
74
+ "817",
75
+ "816",
76
+ "1069",
77
+ "813",
78
+ "812",
79
+ "802",
80
+ "797",
81
+ "796",
82
+ "795",
83
+ "791",
84
+ "790",
85
+ "789",
86
+ "788",
87
+ "786",
88
+ "785",
89
+ "784",
90
+ "783",
91
+ "781",
92
+ "777",
93
+ "776",
94
+ "775",
95
+ "774",
96
+ "770",
97
+ "769",
98
+ "767",
99
+ "764",
100
+ "1063",
101
+ "760",
102
+ "1060",
103
+ "757",
104
+ "756",
105
+ "754",
106
+ "751",
107
+ "749",
108
+ "748",
109
+ "747",
110
+ "746",
111
+ "745",
112
+ "744",
113
+ "742",
114
+ "739",
115
+ "1057",
116
+ "735",
117
+ "734",
118
+ "732",
119
+ "729",
120
+ "1054",
121
+ "725",
122
+ "1051",
123
+ "722",
124
+ "721",
125
+ "719",
126
+ "716",
127
+ "714",
128
+ "713",
129
+ "712",
130
+ "711",
131
+ "710",
132
+ "709",
133
+ "707",
134
+ "704",
135
+ "1048",
136
+ "700",
137
+ "699",
138
+ "697",
139
+ "694",
140
+ "1045",
141
+ "1042",
142
+ "688",
143
+ "687",
144
+ "685",
145
+ "682",
146
+ "680",
147
+ "679",
148
+ "678",
149
+ "677",
150
+ "676",
151
+ "675",
152
+ "673",
153
+ "670",
154
+ "1039",
155
+ "666",
156
+ "665",
157
+ "663",
158
+ "660",
159
+ "1036",
160
+ "656",
161
+ "1033",
162
+ "653",
163
+ "652",
164
+ "650",
165
+ "647",
166
+ "645",
167
+ "644",
168
+ "643",
169
+ "642",
170
+ "641",
171
+ "640",
172
+ "638",
173
+ "635",
174
+ "1030",
175
+ "631",
176
+ "630",
177
+ "628",
178
+ "625",
179
+ "1027",
180
+ "1024",
181
+ "619",
182
+ "618",
183
+ "616",
184
+ "613",
185
+ "611",
186
+ "610",
187
+ "609",
188
+ "608",
189
+ "607",
190
+ "606",
191
+ "604",
192
+ "601",
193
+ "1021",
194
+ "597",
195
+ "596",
196
+ "594",
197
+ "591",
198
+ "1018",
199
+ "587",
200
+ "1015",
201
+ "584",
202
+ "583",
203
+ "581",
204
+ "578",
205
+ "1012",
206
+ "574",
207
+ "573",
208
+ "571",
209
+ "568",
210
+ "1009",
211
+ "564",
212
+ "1006",
213
+ "561",
214
+ "560",
215
+ "558",
216
+ "555",
217
+ "1003",
218
+ "551",
219
+ "550",
220
+ "548",
221
+ "545",
222
+ "1000",
223
+ "541",
224
+ "997",
225
+ "538",
226
+ "537",
227
+ "535",
228
+ "532",
229
+ "994",
230
+ "528",
231
+ "527",
232
+ "525",
233
+ "522",
234
+ "991",
235
+ "988",
236
+ "516",
237
+ "515",
238
+ "513",
239
+ "510",
240
+ "985",
241
+ "506",
242
+ "505",
243
+ "503",
244
+ "500",
245
+ "982",
246
+ "496",
247
+ "979",
248
+ "493",
249
+ "492",
250
+ "490",
251
+ "487",
252
+ "485",
253
+ "484",
254
+ "483",
255
+ "482",
256
+ "481",
257
+ "976",
258
+ "478",
259
+ "973",
260
+ "475",
261
+ "970",
262
+ "472",
263
+ "471",
264
+ "469",
265
+ "466",
266
+ "464",
267
+ "463",
268
+ "462",
269
+ "461",
270
+ "460",
271
+ "967",
272
+ "457",
273
+ "964",
274
+ "961",
275
+ "452",
276
+ "451",
277
+ "449",
278
+ "446",
279
+ "444",
280
+ "443",
281
+ "442",
282
+ "441",
283
+ "440",
284
+ "958",
285
+ "437",
286
+ "955",
287
+ "434",
288
+ "952",
289
+ "431",
290
+ "949",
291
+ "428",
292
+ "946",
293
+ "943",
294
+ "423",
295
+ "940",
296
+ "420",
297
+ "937",
298
+ "417",
299
+ "934",
300
+ "414",
301
+ "931",
302
+ "411",
303
+ "410",
304
+ "408",
305
+ "405",
306
+ "928",
307
+ "401",
308
+ "399",
309
+ "393",
310
+ "392",
311
+ "387",
312
+ "773",
313
+ "1066",
314
+ "794",
315
+ "793",
316
+ "780",
317
+ "779",
318
+ "778",
319
+ "830",
320
+ "829",
321
+ "866",
322
+ "865",
323
+ "892",
324
+ "891",
325
+ "916",
326
+ "911"
327
+ ],
328
+ "constantInitializers": [
329
+ "1001",
330
+ "1002",
331
+ "1004",
332
+ "1005",
333
+ "1007",
334
+ "1008",
335
+ "1010",
336
+ "1011",
337
+ "1013",
338
+ "1014",
339
+ "1016",
340
+ "1017",
341
+ "1019",
342
+ "1020",
343
+ "1022",
344
+ "1023",
345
+ "1025",
346
+ "1026",
347
+ "1028",
348
+ "1029",
349
+ "1031",
350
+ "1032",
351
+ "1034",
352
+ "1035",
353
+ "1037",
354
+ "1038",
355
+ "1040",
356
+ "1041",
357
+ "1043",
358
+ "1044",
359
+ "1046",
360
+ "1047",
361
+ "1049",
362
+ "1050",
363
+ "1052",
364
+ "1053",
365
+ "1055",
366
+ "1056",
367
+ "1058",
368
+ "1059",
369
+ "1061",
370
+ "1062",
371
+ "1064",
372
+ "1065",
373
+ "1067",
374
+ "1068",
375
+ "1070",
376
+ "1071",
377
+ "1073",
378
+ "1074",
379
+ "1076",
380
+ "1077",
381
+ "1079",
382
+ "1080",
383
+ "1082",
384
+ "1083",
385
+ "1086",
386
+ "1090",
387
+ "386",
388
+ "388",
389
+ "389",
390
+ "398",
391
+ "400",
392
+ "752",
393
+ "755",
394
+ "763",
395
+ "801",
396
+ "809",
397
+ "837",
398
+ "845",
399
+ "847",
400
+ "873",
401
+ "890",
402
+ "899",
403
+ "929",
404
+ "930",
405
+ "932",
406
+ "933",
407
+ "935",
408
+ "936",
409
+ "938",
410
+ "939",
411
+ "941",
412
+ "942",
413
+ "944",
414
+ "945",
415
+ "947",
416
+ "948",
417
+ "950",
418
+ "951",
419
+ "953",
420
+ "954",
421
+ "956",
422
+ "957",
423
+ "959",
424
+ "960",
425
+ "962",
426
+ "963",
427
+ "965",
428
+ "966",
429
+ "968",
430
+ "969",
431
+ "971",
432
+ "972",
433
+ "974",
434
+ "975",
435
+ "977",
436
+ "978",
437
+ "980",
438
+ "981",
439
+ "983",
440
+ "984",
441
+ "986",
442
+ "987",
443
+ "989",
444
+ "990",
445
+ "992",
446
+ "993",
447
+ "995",
448
+ "996",
449
+ "998",
450
+ "999",
451
+ "aspp.aspp2.1.weight",
452
+ "backbone.features.11.block.2.fc1.bias",
453
+ "backbone.features.11.block.2.fc1.weight",
454
+ "backbone.features.11.block.2.fc2.bias",
455
+ "backbone.features.11.block.2.fc2.weight",
456
+ "backbone.features.12.block.2.fc1.bias",
457
+ "backbone.features.12.block.2.fc1.weight",
458
+ "backbone.features.12.block.2.fc2.bias",
459
+ "backbone.features.12.block.2.fc2.weight",
460
+ "backbone.features.13.block.2.fc1.bias",
461
+ "backbone.features.13.block.2.fc1.weight",
462
+ "backbone.features.13.block.2.fc2.bias",
463
+ "backbone.features.13.block.2.fc2.weight",
464
+ "backbone.features.14.block.2.fc1.bias",
465
+ "backbone.features.14.block.2.fc1.weight",
466
+ "backbone.features.14.block.2.fc2.bias",
467
+ "backbone.features.14.block.2.fc2.weight",
468
+ "backbone.features.15.block.2.fc1.bias",
469
+ "backbone.features.15.block.2.fc1.weight",
470
+ "backbone.features.15.block.2.fc2.bias",
471
+ "backbone.features.15.block.2.fc2.weight",
472
+ "backbone.features.4.block.2.fc1.bias",
473
+ "backbone.features.4.block.2.fc1.weight",
474
+ "backbone.features.4.block.2.fc2.bias",
475
+ "backbone.features.4.block.2.fc2.weight",
476
+ "backbone.features.5.block.2.fc1.bias",
477
+ "backbone.features.5.block.2.fc1.weight",
478
+ "backbone.features.5.block.2.fc2.bias",
479
+ "backbone.features.5.block.2.fc2.weight",
480
+ "backbone.features.6.block.2.fc1.bias",
481
+ "backbone.features.6.block.2.fc1.weight",
482
+ "backbone.features.6.block.2.fc2.bias",
483
+ "backbone.features.6.block.2.fc2.weight",
484
+ "decoder.decode1.gru.hh.0.bias",
485
+ "decoder.decode1.gru.hh.0.weight",
486
+ "decoder.decode1.gru.ih.0.bias",
487
+ "decoder.decode1.gru.ih.0.weight",
488
+ "decoder.decode2.gru.hh.0.bias",
489
+ "decoder.decode2.gru.hh.0.weight",
490
+ "decoder.decode2.gru.ih.0.bias",
491
+ "decoder.decode2.gru.ih.0.weight",
492
+ "decoder.decode3.gru.hh.0.bias",
493
+ "decoder.decode3.gru.hh.0.weight",
494
+ "decoder.decode3.gru.ih.0.bias",
495
+ "decoder.decode3.gru.ih.0.weight",
496
+ "decoder.decode4.gru.hh.0.bias",
497
+ "decoder.decode4.gru.hh.0.weight",
498
+ "decoder.decode4.gru.ih.0.bias",
499
+ "decoder.decode4.gru.ih.0.weight",
500
+ "project_mat.conv.bias",
501
+ "project_mat.conv.weight"
502
+ ],
503
+ "device": "VAIML",
504
+ "vaimlParam": {
505
+ "vaimlModelPath": "./segmentation_1_4_0_fp32_combined/vaiml_par_0",
506
+ "deviceName": "stx",
507
+ "outputShapes": [
508
+ {
509
+ "shapes": [
510
+ "1",
511
+ "1",
512
+ "180",
513
+ "320"
514
+ ]
515
+ },
516
+ {
517
+ "shapes": [
518
+ "1",
519
+ "16",
520
+ "90",
521
+ "160"
522
+ ]
523
+ },
524
+ {
525
+ "shapes": [
526
+ "1",
527
+ "20",
528
+ "45",
529
+ "80"
530
+ ]
531
+ },
532
+ {
533
+ "shapes": [
534
+ "1",
535
+ "40",
536
+ "23",
537
+ "40"
538
+ ]
539
+ },
540
+ {
541
+ "shapes": [
542
+ "1",
543
+ "64",
544
+ "12",
545
+ "20"
546
+ ]
547
+ },
548
+ {
549
+ "shapes": [
550
+ "1",
551
+ "3",
552
+ "180",
553
+ "320"
554
+ ]
555
+ }
556
+ ],
557
+ "inputShapes": [
558
+ {
559
+ "shapes": [
560
+ "1",
561
+ "180",
562
+ "320",
563
+ "4"
564
+ ]
565
+ },
566
+ {
567
+ "shapes": [
568
+ "1",
569
+ "16",
570
+ "90",
571
+ "160"
572
+ ]
573
+ },
574
+ {
575
+ "shapes": [
576
+ "1",
577
+ "20",
578
+ "45",
579
+ "80"
580
+ ]
581
+ },
582
+ {
583
+ "shapes": [
584
+ "1",
585
+ "40",
586
+ "23",
587
+ "40"
588
+ ]
589
+ },
590
+ {
591
+ "shapes": [
592
+ "1",
593
+ "64",
594
+ "12",
595
+ "20"
596
+ ]
597
+ }
598
+ ],
599
+ "vaimlUnarchivePath": "./segmentation_1_4_0_fp32_combined",
600
+ "subgraphName": "vaiml_par_0",
601
+ "priority": "normal",
602
+ "configJson": "vitisai_config.json",
603
+ "deviceBatchSize": 1,
604
+ "inputNames": [
605
+ "385",
606
+ "394",
607
+ "395",
608
+ "396",
609
+ "397"
610
+ ],
611
+ "outputNames": [
612
+ "921",
613
+ "894",
614
+ "868",
615
+ "832",
616
+ "796",
617
+ "916"
618
+ ]
619
+ }
620
+ }
621
+ ],
622
+ "config": {
623
+ "passes": [
624
+ {
625
+ "name": "init",
626
+ "plugin": "vaip-pass_init"
627
+ },
628
+ {
629
+ "name": "vaiml_partition",
630
+ "plugin": "vaip-pass_vaiml_partition",
631
+ "vaimlConfig": {
632
+ "device": "stx",
633
+ "enableF32ToBf16Conversion": true,
634
+ "keepOutputs": true
635
+ }
636
+ }
637
+ ],
638
+ "cacheDir": "/tmp/vaip/.cache",
639
+ "cacheKey": "4e8d9ee8719e050537726a1cd8f79305",
640
+ "version": {
641
+ "versionInfos": [
642
+ {
643
+ "packageName": "vaip",
644
+ "commit": "e4ca074a034e568a2fd44af176f112e64e845411",
645
+ "version": "vaip.1.0.0"
646
+ },
647
+ {
648
+ "packageName": "target_factory",
649
+ "commit": "60f0780e75861ed37681e00c5eeeaf132f887c45",
650
+ "version": "target-factory.3.5.0"
651
+ },
652
+ {
653
+ "packageName": "vart",
654
+ "commit": "01141f8d14af4be8a1229d3d5dd93f26ba608f58",
655
+ "version": "vart"
656
+ },
657
+ {
658
+ "packageName": "xcompiler",
659
+ "commit": "8ccbd710317a738251bbddcebb919fd0e8145250",
660
+ "version": "xcompiler.3.5.0"
661
+ },
662
+ {
663
+ "packageName": "onnxrutnime",
664
+ "commit": "5c1b7ccbff7e5141c1da7a9d963d660e5741c319",
665
+ "version": "onnxruntime.1.20.1"
666
+ },
667
+ {
668
+ "packageName": "xir",
669
+ "commit": "bef5d269fc06f065c5f8de5ba202d6f02edfad7f",
670
+ "version": "xir.3.5.0"
671
+ },
672
+ {
673
+ "packageName": "xrt",
674
+ "commit": "acc144998d650acbfda7e5919a1290de8f8c7735",
675
+ "version": "xrt.2.19.72"
676
+ },
677
+ {
678
+ "packageName": "graph_engine",
679
+ "commit": "d7385f8afa5f6eb7d5d4452749b0e07bce67b75b",
680
+ "version": "graph_engine"
681
+ }
682
+ ]
683
+ },
684
+ "onnxPath": "segmentation_1_4_0_fp32_combined.onnx",
685
+ "sessionOptions": {
686
+ "config_file": "vitisai_config.json"
687
+ },
688
+ "enableCacheFileIoInMem": false
689
+ },
690
+ "events": [
691
+ {
692
+ "name": "before_compile_onnx_model_internal",
693
+ "ph": "X",
694
+ "ts": "31561",
695
+ "pid": "214",
696
+ "tid": "214",
697
+ "args": {
698
+ "memUsage": {}
699
+ },
700
+ "dur": "30664"
701
+ },
702
+ {
703
+ "id": "before_compile_onnx_model_internal_mem_usage_1",
704
+ "ph": "v",
705
+ "ts": "31561",
706
+ "pid": "214",
707
+ "args": {
708
+ "dumps": {
709
+ "process_totals": {
710
+ "peak_memory": "0",
711
+ "current_memory": "0"
712
+ }
713
+ }
714
+ }
715
+ },
716
+ {
717
+ "id": "before_compile_onnx_model_internal_mem_usage_2",
718
+ "ph": "v",
719
+ "ts": "62226",
720
+ "pid": "214",
721
+ "args": {
722
+ "dumps": {
723
+ "process_totals": {
724
+ "peak_memory": "0",
725
+ "current_memory": "0"
726
+ }
727
+ }
728
+ }
729
+ },
730
+ {
731
+ "name": "check_cache_hit",
732
+ "ph": "X",
733
+ "ts": "62248",
734
+ "pid": "214",
735
+ "tid": "214",
736
+ "args": {
737
+ "memUsage": {}
738
+ },
739
+ "dur": "25"
740
+ },
741
+ {
742
+ "id": "check_cache_hit_mem_usage_1",
743
+ "ph": "v",
744
+ "ts": "62248",
745
+ "pid": "214",
746
+ "args": {
747
+ "dumps": {
748
+ "process_totals": {
749
+ "peak_memory": "0",
750
+ "current_memory": "0"
751
+ }
752
+ }
753
+ }
754
+ },
755
+ {
756
+ "id": "check_cache_hit_mem_usage_2",
757
+ "ph": "v",
758
+ "ts": "62273",
759
+ "pid": "214",
760
+ "args": {
761
+ "dumps": {
762
+ "process_totals": {
763
+ "peak_memory": "0",
764
+ "current_memory": "0"
765
+ }
766
+ }
767
+ }
768
+ },
769
+ {
770
+ "name": "0-init@vaip-pass_init",
771
+ "ph": "X",
772
+ "ts": "65069",
773
+ "pid": "214",
774
+ "tid": "214",
775
+ "args": {
776
+ "memUsage": {}
777
+ },
778
+ "dur": "11"
779
+ },
780
+ {
781
+ "id": "0-init@vaip-pass_init_mem_usage_1",
782
+ "ph": "v",
783
+ "ts": "65069",
784
+ "pid": "214",
785
+ "args": {
786
+ "dumps": {
787
+ "process_totals": {
788
+ "peak_memory": "0",
789
+ "current_memory": "0"
790
+ }
791
+ }
792
+ }
793
+ },
794
+ {
795
+ "id": "0-init@vaip-pass_init_mem_usage_2",
796
+ "ph": "v",
797
+ "ts": "65080",
798
+ "pid": "214",
799
+ "args": {
800
+ "dumps": {
801
+ "process_totals": {
802
+ "peak_memory": "0",
803
+ "current_memory": "0"
804
+ }
805
+ }
806
+ }
807
+ },
808
+ {
809
+ "name": "1-vaiml_partition@vaip-pass_vaiml_partition",
810
+ "ph": "X",
811
+ "ts": "65090",
812
+ "pid": "214",
813
+ "tid": "214",
814
+ "args": {
815
+ "memUsage": {}
816
+ },
817
+ "dur": "8266564167"
818
+ },
819
+ {
820
+ "id": "1-vaiml_partition@vaip-pass_vaiml_partition_mem_usage_1",
821
+ "ph": "v",
822
+ "ts": "65090",
823
+ "pid": "214",
824
+ "args": {
825
+ "dumps": {
826
+ "process_totals": {
827
+ "peak_memory": "0",
828
+ "current_memory": "0"
829
+ }
830
+ }
831
+ }
832
+ },
833
+ {
834
+ "id": "1-vaiml_partition@vaip-pass_vaiml_partition_mem_usage_2",
835
+ "ph": "v",
836
+ "ts": "8266629258",
837
+ "pid": "214",
838
+ "args": {
839
+ "dumps": {
840
+ "process_totals": {
841
+ "peak_memory": "0",
842
+ "current_memory": "0"
843
+ }
844
+ }
845
+ }
846
+ },
847
+ {
848
+ "name": "update_cache",
849
+ "ph": "X",
850
+ "ts": "62287",
851
+ "pid": "214",
852
+ "tid": "214",
853
+ "args": {
854
+ "memUsage": {}
855
+ },
856
+ "dur": "8266568873"
857
+ },
858
+ {
859
+ "id": "update_cache_mem_usage_1",
860
+ "ph": "v",
861
+ "ts": "62287",
862
+ "pid": "214",
863
+ "args": {
864
+ "dumps": {
865
+ "process_totals": {
866
+ "peak_memory": "0",
867
+ "current_memory": "0"
868
+ }
869
+ }
870
+ }
871
+ },
872
+ {
873
+ "id": "update_cache_mem_usage_2",
874
+ "ph": "v",
875
+ "ts": "8266631160",
876
+ "pid": "214",
877
+ "args": {
878
+ "dumps": {
879
+ "process_totals": {
880
+ "peak_memory": "0",
881
+ "current_memory": "0"
882
+ }
883
+ }
884
+ }
885
+ },
886
+ {
887
+ "name": "read_cache",
888
+ "ph": "X",
889
+ "ts": "8266632901",
890
+ "pid": "214",
891
+ "tid": "214",
892
+ "args": {
893
+ "memUsage": {}
894
+ },
895
+ "dur": "1033"
896
+ },
897
+ {
898
+ "id": "read_cache_mem_usage_1",
899
+ "ph": "v",
900
+ "ts": "8266632901",
901
+ "pid": "214",
902
+ "args": {
903
+ "dumps": {
904
+ "process_totals": {
905
+ "peak_memory": "0",
906
+ "current_memory": "0"
907
+ }
908
+ }
909
+ }
910
+ },
911
+ {
912
+ "id": "read_cache_mem_usage_2",
913
+ "ph": "v",
914
+ "ts": "8266633935",
915
+ "pid": "214",
916
+ "args": {
917
+ "dumps": {
918
+ "process_totals": {
919
+ "peak_memory": "0",
920
+ "current_memory": "0"
921
+ }
922
+ }
923
+ }
924
+ },
925
+ {
926
+ "name": "after_compile_onnx_model_internal",
927
+ "ph": "X",
928
+ "ts": "8266633946",
929
+ "pid": "214",
930
+ "tid": "214",
931
+ "args": {
932
+ "memUsage": {}
933
+ },
934
+ "dur": "26451"
935
+ },
936
+ {
937
+ "id": "after_compile_onnx_model_internal_mem_usage_1",
938
+ "ph": "v",
939
+ "ts": "8266633946",
940
+ "pid": "214",
941
+ "args": {
942
+ "dumps": {
943
+ "process_totals": {
944
+ "peak_memory": "0",
945
+ "current_memory": "0"
946
+ }
947
+ }
948
+ }
949
+ },
950
+ {
951
+ "id": "after_compile_onnx_model_internal_mem_usage_2",
952
+ "ph": "v",
953
+ "ts": "8266660397",
954
+ "pid": "214",
955
+ "args": {
956
+ "dumps": {
957
+ "process_totals": {
958
+ "peak_memory": "0",
959
+ "current_memory": "0"
960
+ }
961
+ }
962
+ }
963
+ },
964
+ {
965
+ "name": "compile_onnx_model_internal",
966
+ "ph": "X",
967
+ "ts": "31560",
968
+ "pid": "214",
969
+ "tid": "214",
970
+ "args": {
971
+ "memUsage": {}
972
+ },
973
+ "dur": "8266628844"
974
+ },
975
+ {
976
+ "id": "compile_onnx_model_internal_mem_usage_1",
977
+ "ph": "v",
978
+ "ts": "31560",
979
+ "pid": "214",
980
+ "args": {
981
+ "dumps": {
982
+ "process_totals": {
983
+ "peak_memory": "0",
984
+ "current_memory": "0"
985
+ }
986
+ }
987
+ }
988
+ },
989
+ {
990
+ "id": "compile_onnx_model_internal_mem_usage_2",
991
+ "ph": "v",
992
+ "ts": "8266660404",
993
+ "pid": "214",
994
+ "args": {
995
+ "dumps": {
996
+ "process_totals": {
997
+ "peak_memory": "0",
998
+ "current_memory": "0"
999
+ }
1000
+ }
1001
+ }
1002
+ },
1003
+ {
1004
+ "name": "compile_onnx_model_3",
1005
+ "ph": "X",
1006
+ "ts": "31447",
1007
+ "pid": "214",
1008
+ "tid": "214",
1009
+ "args": {
1010
+ "memUsage": {}
1011
+ },
1012
+ "dur": "8266629035"
1013
+ },
1014
+ {
1015
+ "id": "compile_onnx_model_3_mem_usage_1",
1016
+ "ph": "v",
1017
+ "ts": "31447",
1018
+ "pid": "214",
1019
+ "args": {
1020
+ "dumps": {
1021
+ "process_totals": {
1022
+ "peak_memory": "0",
1023
+ "current_memory": "0"
1024
+ }
1025
+ }
1026
+ }
1027
+ },
1028
+ {
1029
+ "id": "compile_onnx_model_3_mem_usage_2",
1030
+ "ph": "v",
1031
+ "ts": "8266660483",
1032
+ "pid": "214",
1033
+ "args": {
1034
+ "dumps": {
1035
+ "process_totals": {
1036
+ "peak_memory": "0",
1037
+ "current_memory": "0"
1038
+ }
1039
+ }
1040
+ }
1041
+ }
1042
+ ],
1043
+ "cpuUsage": [
1044
+ {
1045
+ "avgCpuUtil": 75.2757797,
1046
+ "memPeakWorkingSetSize": 1998.73828
1047
+ }
1048
+ ],
1049
+ "cacheFiles": [
1050
+ "context.json"
1051
+ ]
1052
+ }
segmentation_1_4_0_fp32_combined/final-vaiml-pass-summary.txt ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ --------- Final Summary of VAIML Pass ----------
2
+ VAIP commit: e4ca074a034e568a2fd44af176f112e64e845411
3
+ Model: segmentation_1_4_0_fp32_combined.onnx
4
+ Model signature: 6bbb891ab96ca9362e0e61024cd02778
5
+ Device: stx
6
+ Model data type: float32
7
+ Device data type: bfloat16
8
+ Number of operators in the model: 317
9
+ GOPs of the model: 3.73517
10
+ Number of operators supported by VAIML: 306 (96.530%)
11
+ GOPs supported by VAIML: 3.735 (100.000%)
12
+ Number of subgraphs supported by VAIML: 1
13
+ Number of operators offloaded by VAIML: 306 (96.530%)
14
+ GOPs offloaded by VAIML: 3.735 (100.000%)
15
+ Number of subgraphs offloaded by VAIML: 1
16
+ Number of subgraphs with compilation errors (fall back to CPU): 0
17
+ Number of subgraphs below 2% GOPs threshold (fall back to CPU): 0
18
+ Stats for offloaded subgraphs
19
+ Subgraph vaiml_par_0 stats:
20
+ Operators: 306 (96.530%)
21
+ GOPs : 3.735 (100.000%) OPs: 3,735,173,696
22
+
23
+
segmentation_1_4_0_fp32_combined/gops.csv ADDED
@@ -0,0 +1,318 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Node,OPs,Note
2
+ Add_105,115200
3
+ Add_115,96000
4
+ Add_124,96000
5
+ Add_132,38400
6
+ Add_135,88320
7
+ Add_144,88320
8
+ Add_152,38400
9
+ Add_155,88320
10
+ Add_164,88320
11
+ Add_172,38400
12
+ Add_175,230400
13
+ Add_184,230400
14
+ Add_19,460800
15
+ Add_196,960
16
+ Add_206,322560
17
+ Add_215,322560
18
+ Add_227,1344
19
+ Add_235,53760
20
+ Add_238,322560
21
+ Add_247,322560
22
+ Add_259,1344
23
+ Add_269,460800
24
+ Add_278,460800
25
+ Add_29,460800
26
+ Add_290,1920
27
+ Add_298,76800
28
+ Add_301,460800
29
+ Add_310,460800
30
+ Add_322,1920
31
+ Add_330,76800
32
+ Add_333,460800
33
+ Add_362,30720
34
+ Add_388,73600
35
+ Add_40,172800
36
+ Add_414,144000
37
+ Add_434,460800
38
+ Add_445,345600
39
+ Add_50,144
40
+ Add_67,240
41
+ Add_75,73600
42
+ Add_85,240
43
+ Add_93,73600
44
+ Add_96,441600
45
+ AveragePool_346,0
46
+ AveragePool_347,0
47
+ AveragePool_348,0
48
+ Cast_0,0
49
+ Clip_108,230400
50
+ Clip_118,192000
51
+ Clip_127,192000
52
+ Clip_138,176640
53
+ Clip_147,176640
54
+ Clip_158,176640
55
+ Clip_167,176640
56
+ Clip_178,460800
57
+ Clip_187,460800
58
+ Clip_199,1920
59
+ Clip_209,645120
60
+ Clip_218,645120
61
+ Clip_22,921600
62
+ Clip_230,2688
63
+ Clip_241,645120
64
+ Clip_250,645120
65
+ Clip_262,2688
66
+ Clip_272,921600
67
+ Clip_281,921600
68
+ Clip_293,3840
69
+ Clip_304,921600
70
+ Clip_313,921600
71
+ Clip_325,3840
72
+ Clip_336,921600
73
+ Clip_446,691200
74
+ Clip_447,230400
75
+ Clip_53,288
76
+ Clip_70,480
77
+ Clip_88,480
78
+ Clip_99,883200
79
+ Concat_350,30720
80
+ Concat_355,30720
81
+ Concat_363,30720
82
+ Concat_372,235520
83
+ Concat_376,73600
84
+ Concat_381,73600
85
+ Concat_389,73600
86
+ Concat_398,576000
87
+ Concat_402,144000
88
+ Concat_407,144000
89
+ Concat_415,144000
90
+ Concat_418,1152000
91
+ Concat_422,460800
92
+ Concat_427,460800
93
+ Concat_435,460800
94
+ Concat_438,3686400
95
+ Conv_103,1152000
96
+ Conv_112,9254400
97
+ Conv_113,7776000
98
+ Conv_122,960000
99
+ Conv_131,7718400
100
+ Conv_133,7153920
101
+ Conv_142,883200
102
+ Conv_151,7104000
103
+ Conv_153,7153920
104
+ Conv_162,883200
105
+ Conv_17,12902400
106
+ Conv_171,7104000
107
+ Conv_173,18662400
108
+ Conv_182,2304000
109
+ Conv_192,115440
110
+ Conv_194,116160
111
+ Conv_203,25858560
112
+ Conv_204,36449280
113
+ Conv_213,3225600
114
+ Conv_223,226128
115
+ Conv_225,227136
116
+ Conv_234,36180480
117
+ Conv_236,36449280
118
+ Conv_245,8386560
119
+ Conv_255,226128
120
+ Conv_257,227136
121
+ Conv_26,4608000
122
+ Conv_266,51686400
123
+ Conv_267,74188800
124
+ Conv_276,11980800
125
+ Conv_28,7833600
126
+ Conv_286,461280
127
+ Conv_288,462720
128
+ Conv_297,73804800
129
+ Conv_299,74188800
130
+ Conv_30,31334400
131
+ Conv_308,11980800
132
+ Conv_318,461280
133
+ Conv_32,4608000
134
+ Conv_320,462720
135
+ Conv_329,73804800
136
+ Conv_331,74188800
137
+ Conv_34,11232000
138
+ Conv_340,59043840
139
+ Conv_343,246016
140
+ Conv_35,12960000
141
+ Conv_351,70840320
142
+ Conv_356,35420160
143
+ Conv_37,5184000
144
+ Conv_373,226688000
145
+ Conv_377,106131200
146
+ Conv_382,53065600
147
+ Conv_39,12614400
148
+ Conv_399,277632000
149
+ Conv_403,103968000
150
+ Conv_408,51984000
151
+ Conv_41,12960000
152
+ Conv_419,490291200
153
+ Conv_423,266342400
154
+ Conv_428,133171200
155
+ Conv_43,3444480
156
+ Conv_439,582451200
157
+ Conv_441,267264000
158
+ Conv_443,7833600
159
+ Conv_46,3504
160
+ Conv_48,3600
161
+ Conv_57,5372800
162
+ Conv_58,9052800
163
+ Conv_60,5740800
164
+ Conv_63,7744
165
+ Conv_65,7920
166
+ Conv_74,8905600
167
+ Conv_76,9052800
168
+ Conv_78,5740800
169
+ Conv_81,7744
170
+ Conv_83,7920
171
+ Conv_92,8905600
172
+ Conv_94,18105600
173
+ Div_101,1766400
174
+ Div_110,460800
175
+ Div_120,384000
176
+ Div_129,384000
177
+ Div_140,353280
178
+ Div_149,353280
179
+ Div_16,1382400
180
+ Div_160,353280
181
+ Div_169,353280
182
+ Div_180,921600
183
+ Div_189,921600
184
+ Div_2,1843200
185
+ Div_201,3840
186
+ Div_211,1290240
187
+ Div_220,1290240
188
+ Div_232,5376
189
+ Div_24,1843200
190
+ Div_243,1290240
191
+ Div_252,1290240
192
+ Div_264,5376
193
+ Div_274,1843200
194
+ Div_283,1843200
195
+ Div_295,7680
196
+ Div_306,1843200
197
+ Div_315,1843200
198
+ Div_327,7680
199
+ Div_338,1843200
200
+ Div_55,576
201
+ Div_72,960
202
+ Div_90,960
203
+ GlobalAveragePool_191,234240
204
+ GlobalAveragePool_222,327936
205
+ GlobalAveragePool_254,327936
206
+ GlobalAveragePool_285,468480
207
+ GlobalAveragePool_317,468480
208
+ GlobalAveragePool_342,468480
209
+ GlobalAveragePool_45,133056
210
+ GlobalAveragePool_62,221760
211
+ GlobalAveragePool_80,221760
212
+ Mul_102,441600
213
+ Mul_111,115200
214
+ Mul_121,96000
215
+ Mul_130,96000
216
+ Mul_141,88320
217
+ Mul_150,88320
218
+ Mul_161,88320
219
+ Mul_170,88320
220
+ Mul_181,230400
221
+ Mul_190,230400
222
+ Mul_202,230400
223
+ Mul_212,322560
224
+ Mul_221,322560
225
+ Mul_233,322560
226
+ Mul_244,322560
227
+ Mul_25,460800
228
+ Mul_253,322560
229
+ Mul_265,322560
230
+ Mul_275,460800
231
+ Mul_284,460800
232
+ Mul_296,460800
233
+ Mul_307,460800
234
+ Mul_316,460800
235
+ Mul_328,460800
236
+ Mul_339,460800
237
+ Mul_345,61440
238
+ Mul_354,30720
239
+ Mul_360,30720
240
+ Mul_361,30720
241
+ Mul_380,73600
242
+ Mul_386,73600
243
+ Mul_387,73600
244
+ Mul_406,144000
245
+ Mul_412,144000
246
+ Mul_413,144000
247
+ Mul_426,460800
248
+ Mul_432,460800
249
+ Mul_433,460800
250
+ Mul_56,132480
251
+ Mul_73,220800
252
+ Mul_91,220800
253
+ Relu_193,240
254
+ Relu_224,336
255
+ Relu_256,336
256
+ Relu_27,460800
257
+ Relu_287,480
258
+ Relu_31,1843200
259
+ Relu_319,480
260
+ Relu_33,460800
261
+ Relu_341,61440
262
+ Relu_36,518400
263
+ Relu_374,147200
264
+ Relu_38,518400
265
+ Relu_400,288000
266
+ Relu_42,518400
267
+ Relu_420,921600
268
+ Relu_44,132480
269
+ Relu_440,1843200
270
+ Relu_442,1843200
271
+ Relu_47,48
272
+ Relu_59,220800
273
+ Relu_61,220800
274
+ Relu_64,64
275
+ Relu_77,220800
276
+ Relu_79,220800
277
+ Relu_82,64
278
+ Resize_365,245760
279
+ Resize_391,588800
280
+ Resize_417,1152000
281
+ Resize_437,3686400
282
+ Sigmoid_344,0
283
+ Sigmoid_352,0
284
+ Sigmoid_378,0
285
+ Sigmoid_404,0
286
+ Sigmoid_424,0
287
+ Slice_371,235520
288
+ Slice_397,576000
289
+ Slice_7,345600
290
+ Split_349,0
291
+ Split_353,0
292
+ Split_375,0
293
+ Split_379,0
294
+ Split_401,0
295
+ Split_405,0
296
+ Split_421,0
297
+ Split_425,0
298
+ Split_444,0
299
+ Sub_14,345600
300
+ Sub_359,30720
301
+ Sub_385,73600
302
+ Sub_411,144000
303
+ Sub_431,460800
304
+ Tanh_357,2334720
305
+ Tanh_383,5593600
306
+ Tanh_409,10944000
307
+ Tanh_429,35020800
308
+ Transpose_10,0
309
+ Transpose_11,0
310
+ Transpose_12,0
311
+ Transpose_448,0
312
+ Transpose_449,0
313
+ Transpose_450,0
314
+ Transpose_451,0
315
+ Transpose_452,0
316
+ Transpose_453,0
317
+ Transpose_8,0
318
+ Transpose_9,0
segmentation_1_4_0_fp32_combined/graph_partition_trace.csv ADDED
@@ -0,0 +1,318 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Node, Type, Subgraph/CustomOp, Status
2
+ Add_105,,vaiml_par_0,Supported
3
+ Add_115,,vaiml_par_0,Supported
4
+ Add_124,,vaiml_par_0,Supported
5
+ Add_132,,vaiml_par_0,Supported
6
+ Add_135,,vaiml_par_0,Supported
7
+ Add_144,,vaiml_par_0,Supported
8
+ Add_152,,vaiml_par_0,Supported
9
+ Add_155,,vaiml_par_0,Supported
10
+ Add_164,,vaiml_par_0,Supported
11
+ Add_172,,vaiml_par_0,Supported
12
+ Add_175,,vaiml_par_0,Supported
13
+ Add_184,,vaiml_par_0,Supported
14
+ Add_19,,vaiml_par_0,Supported
15
+ Add_196,,vaiml_par_0,Supported
16
+ Add_206,,vaiml_par_0,Supported
17
+ Add_215,,vaiml_par_0,Supported
18
+ Add_227,,vaiml_par_0,Supported
19
+ Add_235,,vaiml_par_0,Supported
20
+ Add_238,,vaiml_par_0,Supported
21
+ Add_247,,vaiml_par_0,Supported
22
+ Add_259,,vaiml_par_0,Supported
23
+ Add_269,,vaiml_par_0,Supported
24
+ Add_278,,vaiml_par_0,Supported
25
+ Add_29,,vaiml_par_0,Supported
26
+ Add_290,,vaiml_par_0,Supported
27
+ Add_298,,vaiml_par_0,Supported
28
+ Add_301,,vaiml_par_0,Supported
29
+ Add_310,,vaiml_par_0,Supported
30
+ Add_322,,vaiml_par_0,Supported
31
+ Add_330,,vaiml_par_0,Supported
32
+ Add_333,,vaiml_par_0,Supported
33
+ Add_362,,vaiml_par_0,Supported
34
+ Add_388,,vaiml_par_0,Supported
35
+ Add_40,,vaiml_par_0,Supported
36
+ Add_414,,vaiml_par_0,Supported
37
+ Add_434,,vaiml_par_0,Supported
38
+ Add_445,,vaiml_par_0,Supported
39
+ Add_50,,vaiml_par_0,Supported
40
+ Add_67,,vaiml_par_0,Supported
41
+ Add_75,,vaiml_par_0,Supported
42
+ Add_85,,vaiml_par_0,Supported
43
+ Add_93,,vaiml_par_0,Supported
44
+ Add_96,,vaiml_par_0,Supported
45
+ AveragePool_346,,vaiml_par_0,Supported
46
+ AveragePool_347,,vaiml_par_0,Supported
47
+ AveragePool_348,,vaiml_par_0,Supported
48
+ Cast_0,,,Not supported. Check aie_unsupported_original_ops.json
49
+ Clip_108,,vaiml_par_0,Supported
50
+ Clip_118,,vaiml_par_0,Supported
51
+ Clip_127,,vaiml_par_0,Supported
52
+ Clip_138,,vaiml_par_0,Supported
53
+ Clip_147,,vaiml_par_0,Supported
54
+ Clip_158,,vaiml_par_0,Supported
55
+ Clip_167,,vaiml_par_0,Supported
56
+ Clip_178,,vaiml_par_0,Supported
57
+ Clip_187,,vaiml_par_0,Supported
58
+ Clip_199,,vaiml_par_0,Supported
59
+ Clip_209,,vaiml_par_0,Supported
60
+ Clip_218,,vaiml_par_0,Supported
61
+ Clip_22,,vaiml_par_0,Supported
62
+ Clip_230,,vaiml_par_0,Supported
63
+ Clip_241,,vaiml_par_0,Supported
64
+ Clip_250,,vaiml_par_0,Supported
65
+ Clip_262,,vaiml_par_0,Supported
66
+ Clip_272,,vaiml_par_0,Supported
67
+ Clip_281,,vaiml_par_0,Supported
68
+ Clip_293,,vaiml_par_0,Supported
69
+ Clip_304,,vaiml_par_0,Supported
70
+ Clip_313,,vaiml_par_0,Supported
71
+ Clip_325,,vaiml_par_0,Supported
72
+ Clip_336,,vaiml_par_0,Supported
73
+ Clip_446,,vaiml_par_0,Supported
74
+ Clip_447,,vaiml_par_0,Supported
75
+ Clip_53,,vaiml_par_0,Supported
76
+ Clip_70,,vaiml_par_0,Supported
77
+ Clip_88,,vaiml_par_0,Supported
78
+ Clip_99,,vaiml_par_0,Supported
79
+ Concat_350,,vaiml_par_0,Supported
80
+ Concat_355,,vaiml_par_0,Supported
81
+ Concat_363,,vaiml_par_0,Supported
82
+ Concat_372,,vaiml_par_0,Supported
83
+ Concat_376,,vaiml_par_0,Supported
84
+ Concat_381,,vaiml_par_0,Supported
85
+ Concat_389,,vaiml_par_0,Supported
86
+ Concat_398,,vaiml_par_0,Supported
87
+ Concat_402,,vaiml_par_0,Supported
88
+ Concat_407,,vaiml_par_0,Supported
89
+ Concat_415,,vaiml_par_0,Supported
90
+ Concat_418,,vaiml_par_0,Supported
91
+ Concat_422,,vaiml_par_0,Supported
92
+ Concat_427,,vaiml_par_0,Supported
93
+ Concat_435,,vaiml_par_0,Supported
94
+ Concat_438,,vaiml_par_0,Supported
95
+ Conv_103,,vaiml_par_0,Supported
96
+ Conv_112,,vaiml_par_0,Supported
97
+ Conv_113,,vaiml_par_0,Supported
98
+ Conv_122,,vaiml_par_0,Supported
99
+ Conv_131,,vaiml_par_0,Supported
100
+ Conv_133,,vaiml_par_0,Supported
101
+ Conv_142,,vaiml_par_0,Supported
102
+ Conv_151,,vaiml_par_0,Supported
103
+ Conv_153,,vaiml_par_0,Supported
104
+ Conv_162,,vaiml_par_0,Supported
105
+ Conv_17,,vaiml_par_0,Supported
106
+ Conv_171,,vaiml_par_0,Supported
107
+ Conv_173,,vaiml_par_0,Supported
108
+ Conv_182,,vaiml_par_0,Supported
109
+ Conv_192,,vaiml_par_0,Supported
110
+ Conv_194,,vaiml_par_0,Supported
111
+ Conv_203,,vaiml_par_0,Supported
112
+ Conv_204,,vaiml_par_0,Supported
113
+ Conv_213,,vaiml_par_0,Supported
114
+ Conv_223,,vaiml_par_0,Supported
115
+ Conv_225,,vaiml_par_0,Supported
116
+ Conv_234,,vaiml_par_0,Supported
117
+ Conv_236,,vaiml_par_0,Supported
118
+ Conv_245,,vaiml_par_0,Supported
119
+ Conv_255,,vaiml_par_0,Supported
120
+ Conv_257,,vaiml_par_0,Supported
121
+ Conv_26,,vaiml_par_0,Supported
122
+ Conv_266,,vaiml_par_0,Supported
123
+ Conv_267,,vaiml_par_0,Supported
124
+ Conv_276,,vaiml_par_0,Supported
125
+ Conv_28,,vaiml_par_0,Supported
126
+ Conv_286,,vaiml_par_0,Supported
127
+ Conv_288,,vaiml_par_0,Supported
128
+ Conv_297,,vaiml_par_0,Supported
129
+ Conv_299,,vaiml_par_0,Supported
130
+ Conv_30,,vaiml_par_0,Supported
131
+ Conv_308,,vaiml_par_0,Supported
132
+ Conv_318,,vaiml_par_0,Supported
133
+ Conv_32,,vaiml_par_0,Supported
134
+ Conv_320,,vaiml_par_0,Supported
135
+ Conv_329,,vaiml_par_0,Supported
136
+ Conv_331,,vaiml_par_0,Supported
137
+ Conv_34,,vaiml_par_0,Supported
138
+ Conv_340,,vaiml_par_0,Supported
139
+ Conv_343,,vaiml_par_0,Supported
140
+ Conv_35,,vaiml_par_0,Supported
141
+ Conv_351,,vaiml_par_0,Supported
142
+ Conv_356,,vaiml_par_0,Supported
143
+ Conv_37,,vaiml_par_0,Supported
144
+ Conv_373,,vaiml_par_0,Supported
145
+ Conv_377,,vaiml_par_0,Supported
146
+ Conv_382,,vaiml_par_0,Supported
147
+ Conv_39,,vaiml_par_0,Supported
148
+ Conv_399,,vaiml_par_0,Supported
149
+ Conv_403,,vaiml_par_0,Supported
150
+ Conv_408,,vaiml_par_0,Supported
151
+ Conv_41,,vaiml_par_0,Supported
152
+ Conv_419,,vaiml_par_0,Supported
153
+ Conv_423,,vaiml_par_0,Supported
154
+ Conv_428,,vaiml_par_0,Supported
155
+ Conv_43,,vaiml_par_0,Supported
156
+ Conv_439,,vaiml_par_0,Supported
157
+ Conv_441,,vaiml_par_0,Supported
158
+ Conv_443,,vaiml_par_0,Supported
159
+ Conv_46,,vaiml_par_0,Supported
160
+ Conv_48,,vaiml_par_0,Supported
161
+ Conv_57,,vaiml_par_0,Supported
162
+ Conv_58,,vaiml_par_0,Supported
163
+ Conv_60,,vaiml_par_0,Supported
164
+ Conv_63,,vaiml_par_0,Supported
165
+ Conv_65,,vaiml_par_0,Supported
166
+ Conv_74,,vaiml_par_0,Supported
167
+ Conv_76,,vaiml_par_0,Supported
168
+ Conv_78,,vaiml_par_0,Supported
169
+ Conv_81,,vaiml_par_0,Supported
170
+ Conv_83,,vaiml_par_0,Supported
171
+ Conv_92,,vaiml_par_0,Supported
172
+ Conv_94,,vaiml_par_0,Supported
173
+ Div_101,,vaiml_par_0,Supported
174
+ Div_110,,vaiml_par_0,Supported
175
+ Div_120,,vaiml_par_0,Supported
176
+ Div_129,,vaiml_par_0,Supported
177
+ Div_140,,vaiml_par_0,Supported
178
+ Div_149,,vaiml_par_0,Supported
179
+ Div_16,,vaiml_par_0,Supported
180
+ Div_160,,vaiml_par_0,Supported
181
+ Div_169,,vaiml_par_0,Supported
182
+ Div_180,,vaiml_par_0,Supported
183
+ Div_189,,vaiml_par_0,Supported
184
+ Div_2,,vaiml_par_0,Supported
185
+ Div_201,,vaiml_par_0,Supported
186
+ Div_211,,vaiml_par_0,Supported
187
+ Div_220,,vaiml_par_0,Supported
188
+ Div_232,,vaiml_par_0,Supported
189
+ Div_24,,vaiml_par_0,Supported
190
+ Div_243,,vaiml_par_0,Supported
191
+ Div_252,,vaiml_par_0,Supported
192
+ Div_264,,vaiml_par_0,Supported
193
+ Div_274,,vaiml_par_0,Supported
194
+ Div_283,,vaiml_par_0,Supported
195
+ Div_295,,vaiml_par_0,Supported
196
+ Div_306,,vaiml_par_0,Supported
197
+ Div_315,,vaiml_par_0,Supported
198
+ Div_327,,vaiml_par_0,Supported
199
+ Div_338,,vaiml_par_0,Supported
200
+ Div_55,,vaiml_par_0,Supported
201
+ Div_72,,vaiml_par_0,Supported
202
+ Div_90,,vaiml_par_0,Supported
203
+ GlobalAveragePool_191,,vaiml_par_0,Supported
204
+ GlobalAveragePool_222,,vaiml_par_0,Supported
205
+ GlobalAveragePool_254,,vaiml_par_0,Supported
206
+ GlobalAveragePool_285,,vaiml_par_0,Supported
207
+ GlobalAveragePool_317,,vaiml_par_0,Supported
208
+ GlobalAveragePool_342,,vaiml_par_0,Supported
209
+ GlobalAveragePool_45,,vaiml_par_0,Supported
210
+ GlobalAveragePool_62,,vaiml_par_0,Supported
211
+ GlobalAveragePool_80,,vaiml_par_0,Supported
212
+ Mul_102,,vaiml_par_0,Supported
213
+ Mul_111,,vaiml_par_0,Supported
214
+ Mul_121,,vaiml_par_0,Supported
215
+ Mul_130,,vaiml_par_0,Supported
216
+ Mul_141,,vaiml_par_0,Supported
217
+ Mul_150,,vaiml_par_0,Supported
218
+ Mul_161,,vaiml_par_0,Supported
219
+ Mul_170,,vaiml_par_0,Supported
220
+ Mul_181,,vaiml_par_0,Supported
221
+ Mul_190,,vaiml_par_0,Supported
222
+ Mul_202,,vaiml_par_0,Supported
223
+ Mul_212,,vaiml_par_0,Supported
224
+ Mul_221,,vaiml_par_0,Supported
225
+ Mul_233,,vaiml_par_0,Supported
226
+ Mul_244,,vaiml_par_0,Supported
227
+ Mul_25,,vaiml_par_0,Supported
228
+ Mul_253,,vaiml_par_0,Supported
229
+ Mul_265,,vaiml_par_0,Supported
230
+ Mul_275,,vaiml_par_0,Supported
231
+ Mul_284,,vaiml_par_0,Supported
232
+ Mul_296,,vaiml_par_0,Supported
233
+ Mul_307,,vaiml_par_0,Supported
234
+ Mul_316,,vaiml_par_0,Supported
235
+ Mul_328,,vaiml_par_0,Supported
236
+ Mul_339,,vaiml_par_0,Supported
237
+ Mul_345,,vaiml_par_0,Supported
238
+ Mul_354,,vaiml_par_0,Supported
239
+ Mul_360,,vaiml_par_0,Supported
240
+ Mul_361,,vaiml_par_0,Supported
241
+ Mul_380,,vaiml_par_0,Supported
242
+ Mul_386,,vaiml_par_0,Supported
243
+ Mul_387,,vaiml_par_0,Supported
244
+ Mul_406,,vaiml_par_0,Supported
245
+ Mul_412,,vaiml_par_0,Supported
246
+ Mul_413,,vaiml_par_0,Supported
247
+ Mul_426,,vaiml_par_0,Supported
248
+ Mul_432,,vaiml_par_0,Supported
249
+ Mul_433,,vaiml_par_0,Supported
250
+ Mul_56,,vaiml_par_0,Supported
251
+ Mul_73,,vaiml_par_0,Supported
252
+ Mul_91,,vaiml_par_0,Supported
253
+ Relu_193,,vaiml_par_0,Supported
254
+ Relu_224,,vaiml_par_0,Supported
255
+ Relu_256,,vaiml_par_0,Supported
256
+ Relu_27,,vaiml_par_0,Supported
257
+ Relu_287,,vaiml_par_0,Supported
258
+ Relu_31,,vaiml_par_0,Supported
259
+ Relu_319,,vaiml_par_0,Supported
260
+ Relu_33,,vaiml_par_0,Supported
261
+ Relu_341,,vaiml_par_0,Supported
262
+ Relu_36,,vaiml_par_0,Supported
263
+ Relu_374,,vaiml_par_0,Supported
264
+ Relu_38,,vaiml_par_0,Supported
265
+ Relu_400,,vaiml_par_0,Supported
266
+ Relu_42,,vaiml_par_0,Supported
267
+ Relu_420,,vaiml_par_0,Supported
268
+ Relu_44,,vaiml_par_0,Supported
269
+ Relu_440,,vaiml_par_0,Supported
270
+ Relu_442,,vaiml_par_0,Supported
271
+ Relu_47,,vaiml_par_0,Supported
272
+ Relu_59,,vaiml_par_0,Supported
273
+ Relu_61,,vaiml_par_0,Supported
274
+ Relu_64,,vaiml_par_0,Supported
275
+ Relu_77,,vaiml_par_0,Supported
276
+ Relu_79,,vaiml_par_0,Supported
277
+ Relu_82,,vaiml_par_0,Supported
278
+ Resize_365,,vaiml_par_0,Supported
279
+ Resize_391,,vaiml_par_0,Supported
280
+ Resize_417,,vaiml_par_0,Supported
281
+ Resize_437,,vaiml_par_0,Supported
282
+ Sigmoid_344,,vaiml_par_0,Supported
283
+ Sigmoid_352,,vaiml_par_0,Supported
284
+ Sigmoid_378,,vaiml_par_0,Supported
285
+ Sigmoid_404,,vaiml_par_0,Supported
286
+ Sigmoid_424,,vaiml_par_0,Supported
287
+ Slice_371,,vaiml_par_0,Supported
288
+ Slice_397,,vaiml_par_0,Supported
289
+ Slice_7,,vaiml_par_0,Supported
290
+ Split_349,,vaiml_par_0,Supported
291
+ Split_353,,vaiml_par_0,Supported
292
+ Split_375,,vaiml_par_0,Supported
293
+ Split_379,,vaiml_par_0,Supported
294
+ Split_401,,vaiml_par_0,Supported
295
+ Split_405,,vaiml_par_0,Supported
296
+ Split_421,,vaiml_par_0,Supported
297
+ Split_425,,vaiml_par_0,Supported
298
+ Split_444,,vaiml_par_0,Supported
299
+ Sub_14,,vaiml_par_0,Supported
300
+ Sub_359,,vaiml_par_0,Supported
301
+ Sub_385,,vaiml_par_0,Supported
302
+ Sub_411,,vaiml_par_0,Supported
303
+ Sub_431,,vaiml_par_0,Supported
304
+ Tanh_357,,vaiml_par_0,Supported
305
+ Tanh_383,,vaiml_par_0,Supported
306
+ Tanh_409,,vaiml_par_0,Supported
307
+ Tanh_429,,vaiml_par_0,Supported
308
+ Transpose_10,,,Not supported. Check aie_unsupported_original_ops.json
309
+ Transpose_11,,,Not supported. Check aie_unsupported_original_ops.json
310
+ Transpose_12,,,Not supported. Check aie_unsupported_original_ops.json
311
+ Transpose_448,,,Not supported. Check aie_unsupported_original_ops.json
312
+ Transpose_449,,,Not supported. Check aie_unsupported_original_ops.json
313
+ Transpose_450,,,Not supported. Check aie_unsupported_original_ops.json
314
+ Transpose_451,,,Not supported. Check aie_unsupported_original_ops.json
315
+ Transpose_452,,,Not supported. Check aie_unsupported_original_ops.json
316
+ Transpose_453,,,Not supported. Check aie_unsupported_original_ops.json
317
+ Transpose_8,,vaiml_par_0,Supported
318
+ Transpose_9,,,Not supported. Check aie_unsupported_original_ops.json
segmentation_1_4_0_fp32_combined/original-info-signature.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ 4f00fb244983f7c2158dc9333522f122
segmentation_1_4_0_fp32_combined/original-model-signature.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ 6bbb891ab96ca9362e0e61024cd02778
segmentation_1_4_0_fp32_combined/preliminary-vaiml-pass-summary.txt ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ----- Preliminary Summary of VAIML Pass ------
2
+ Model data type: float32
3
+ Device data type: bfloat16
4
+ Number of operators in the model: 317
5
+ GOPs of the model: 3.73517
6
+ Number of operators supported by VAIML: 306(96.530%)
7
+ GOPs supported by VAIML: 3.735 (100.000%)
8
+ Number of subgraphs supported by VAIML: 1
9
+
10
+ Number of identified subgraphs: 1
11
+ Stats for identified subgraphs
12
+ vaiml_par_0 stats:
13
+ Operators: 306 (96.530%)
14
+ GOPs: 3.735(100.000%) OPs: 3,735,173,696
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/AIECompiler.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/.target ADDED
@@ -0,0 +1 @@
 
 
1
+ hw
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/0_0.log ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
2
+ Configuration: Release_LLVM
3
+ Compiling "0_0.ll"
4
+ chess-clang --chess-proc-dir=/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -S -O2 -std=c++2a -fno-builtin-memcpy -mllvm -instcombine-code-sinking=false -mllvm -disable-lsr -mllvm -replexitval=never -mllvm -enable-load-pre=false -mllvm -chess-disable-add-to-or -mllvm -chess-combine-gep-indices=none -mllvm -chess-disable-fold-phi-of-loads -mllvm -chess-aainfo2chains-algo=4 -mllvm -chess-aggressive-aainfo=false -mllvm -chess-enable-indvarsimplify=0 -mllvm -chess-disable-cse-across-loopboundary -mllvm -chess-tbaa-detect-common-underlying-object=true -mllvm -chess-protect-llvm-global-reg-access=true -fno-jump-tables -fno-discard-value-names -g ../../ir/0_0.ll -o../Release/chesswork848/0_0.sfg --chess-proc-name=me
5
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
6
+ noodle -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -iaie_core.h +Sinl +Olbb=200 +Opmsa +NOpld +Olzyinl +w../Release/chesswork848 ../Release/chesswork848/0_0.sfg +Q1=+Sinl,+Olbb=200,+Opmsa,+NOpld,+Olzyinl +Q2=+Sinl,+Olbb=200,+Opmsa,+NOpld,+Olzyinl +Q3=+Sinl,+Olbb=1000,+Opmsa,+NOpld,+Olzyinl +Qfast=+Sinl,+Olbb=1000,+Opmsa,+NOpld,+Olzyinl,+Opfp +Qs=+Sinl,+Olbb=200,+Opmsa,+NOpld,+Olzyinl +Qz=+Sinl,+Olbb=200,+Opmsa,+NOpld,+Olzyinl me
7
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
8
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
9
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
10
+ chess-backend 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation --print-subtools --cosel -m +ef +s -M3 --amnesia -p10 -q2 -ecrSCDEn -ecrMCDEn -ecrVaddSign -ecrUnpackSign -ecrPackSign -ecrUPSSign -ecrUPSMode -ecrSRSSign -ecrSRSMode -ecrF2IMask -ecrUnpackSize -ecrPackSize -ecrSat -ecrRnd +Oefc +Opbr +Odhls +Oprefer-local-reg-moves -Onocb --mist1 -k64 --showcolor -b -Obbl --mist2 -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 +Omod -k64 +Omsbr=100 +Opnll +A +pnopALU +pnopLDA +pnopLDB +pnopSTS +pnopVEC +Ofexm +Omsanafs +Onzmem +Onombt +Ochex +Omsmfi +Omslactc=lckLdaRsrc_E1,lckLdbRsrc_E1:2 +Odra +Oslr=crSRSSign +Oslr=crUPSMode +Oslr=crSRSMode +Oslr=crRnd +Oslr=crSat +Onop-syntax=NOPA +Onop-syntax=NOPB +Onop-syntax=NOPM +Onop-syntax=NOPV +Onop-syntax=NOPS +Onop-syntax=NOPX +Onop-syntax=NOPXM --tale -g -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -V0_0 -L
11
+ chess-backend 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x --print-subtools --cosel -m +ef +s -M3 --amnesia -p10 -q2 -ecrSCDEn -ecrMCDEn -ecrVaddSign -ecrUnpackSign -ecrPackSign -ecrUPSSign -ecrUPSMode -ecrSRSSign -ecrSRSMode -ecrF2IMask -ecrUnpackSize -ecrPackSize -ecrSat -ecrRnd +Oefc +Opbr +Odhls +Oprefer-local-reg-moves -Onocb --mist1 -k64 --showcolor -b -Obbl --mist2 -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 +Omod -k64 +Omsbr=100 +Opnll +A +pnopALU +pnopLDA +pnopLDB +pnopSTS +pnopVEC +Ofexm +Omsanafs +Onzmem +Onombt +Ochex +Omsmfi +Omslactc=lckLdaRsrc_E1,lckLdbRsrc_E1:2 +Odra +Oslr=crSRSSign +Oslr=crUPSMode +Oslr=crSRSMode +Oslr=crRnd +Oslr=crSat +Onop-syntax=NOPA +Onop-syntax=NOPB +Onop-syntax=NOPM +Onop-syntax=NOPV +Onop-syntax=NOPS +Onop-syntax=NOPX +Onop-syntax=NOPXM --tale -g -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -V0_0 -L
12
+ chess-backend --gvt me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation --tale -g -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -V0_0 -L
13
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
14
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
15
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
16
+ --cosel -m +ef +s -M3 --common 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation
17
+ --cosel -m +ef +s -M3 --common 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x
18
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
19
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
20
+ --amnesia -p10 -q2 -ecrSCDEn -ecrMCDEn -ecrVaddSign -ecrUnpackSign -ecrPackSign -ecrUPSSign -ecrUPSMode -ecrSRSSign -ecrSRSMode -ecrF2IMask -ecrUnpackSize -ecrPackSize -ecrSat -ecrRnd +Oefc +Opbr +Odhls +Oprefer-local-reg-moves -Onocb --common 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation
21
+ --mist1 -k64 --common 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation
22
+ --amnesia -p10 -q2 -ecrSCDEn -ecrMCDEn -ecrVaddSign -ecrUnpackSign -ecrPackSign -ecrUPSSign -ecrUPSMode -ecrSRSSign -ecrSRSMode -ecrF2IMask -ecrUnpackSize -ecrPackSize -ecrSat -ecrRnd +Oefc +Opbr +Odhls +Oprefer-local-reg-moves -Onocb --common 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x
23
+ --showcolor -b -Obbl --common 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation
24
+ --mist2 -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 +Omod -k64 +Omsbr=100 +Opnll +A +pnopALU +pnopLDA +pnopLDB +pnopSTS +pnopVEC +Ofexm +Omsanafs +Onzmem +Onombt +Ochex +Omsmfi +Omslactc=lckLdaRsrc_E1,lckLdbRsrc_E1:2 +Odra +Oslr=crSRSSign +Oslr=crUPSMode +Oslr=crSRSMode +Oslr=crRnd +Oslr=crSat +Onop-syntax=NOPA +Onop-syntax=NOPB +Onop-syntax=NOPM +Onop-syntax=NOPV +Onop-syntax=NOPS +Onop-syntax=NOPX +Onop-syntax=NOPXM --common 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation
25
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
26
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
27
+ Warning in "/usr/local/lib/python3.10/dist-packages/include/adf/aie/tile_control.h", line 292, column 8: in "/usr/local/lib/python3.10/dist-packages/include/adf/aie/tile_control.h", line 292: (loop #8)
28
+ loop software pipelining (to 2 cycles) is feasible for a minimum loop count of 5,
29
+ but requires the creation of a post-amble, for which the loop was not prepared
30
+ ... consider annotating the loop with `chess_prepare_for_pipelining', as well as
31
+ increasing the current `chess_loop_range(1,)` annotation to `chess_loop_range(5,)', or remove it.
32
+
33
+ --tale -g -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -V0_0 -L --common 0_0-F_ZN3adf11block_writeEPKNS_7reg_valEj_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation
34
+ --mist1 -k64 --common 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x
35
+ --showcolor -b -Obbl --common 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x
36
+ --mist2 -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 +Omod -k64 +Omsbr=100 +Opnll +A +pnopALU +pnopLDA +pnopLDB +pnopSTS +pnopVEC +Ofexm +Omsanafs +Onzmem +Onombt +Ochex +Omsmfi +Omslactc=lckLdaRsrc_E1,lckLdbRsrc_E1:2 +Odra +Oslr=crSRSSign +Oslr=crUPSMode +Oslr=crSRSMode +Oslr=crRnd +Oslr=crSat +Onop-syntax=NOPA +Onop-syntax=NOPB +Onop-syntax=NOPM +Onop-syntax=NOPV +Onop-syntax=NOPS +Onop-syntax=NOPX +Onop-syntax=NOPXM --common 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x
37
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
38
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
39
+ Warning in "0_0/src/0_0.cc", line 37, column 12: in "0_0/src/0_0.cc", line 37: (loop #13)
40
+ loop software pipelining (to 8 cycles) is feasible but requires the creation of a post-amble,
41
+ for which the loop was not prepared
42
+ ... consider annotating the loop with `chess_prepare_for_pipelining'
43
+
44
+ Warning: in "0_0/src/0_0.cc", line 12: (loop #3)
45
+ Non leaf loop was prepared for pipelining. But the pipelined solutions have not been selected.
46
+ Consider removing the chess_prepare_for_pipelining directive as it may improve results
47
+ --tale -g -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -V0_0 -L --common 0_0-main_ me /usr/local/lib/python3.10/dist-packages/data/aie2p/lib +H/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/elongation -x
48
+ bridge -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -i -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 0_0.objlist -o../0_0.o -pme
49
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
50
+ darts -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -d -h -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -L +Ihex +nanno ../Release/0_0.o me
51
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
52
+ Linking "../Release/0_0"
53
+ bridge -o../Release/0_0 ../Release/0_0.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork848 -pme
54
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
55
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
56
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
57
+ darts -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -d -h -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -L +Ihex +nanno +u ../Release/0_0 me
58
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
59
+ Compilation finished successfully (19 errors, 3 warnings)
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0 ADDED
Binary file (47.4 kB). View file
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.# ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ 3d25a3bfdb1b94e31ca421fe169265bb6b32498c
2
+ 1a735f496f1284ab86135ede88939a8aeeb375a3
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.## ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ 46f16ef751fb28f6dce00ef4d676df021c165b9a
2
+ 92e103875c6a39bc8cdbe4a21c02c5a94ba3cc96
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.calltree ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:42:51 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0 ../Release/0_0.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork848 -pme
5
+
6
+
7
+ // Release: ipp V-2024.06-TGT-241219
8
+
9
+ _main_init
10
+ _main
11
+ _ZN3adf11block_writeEPKNS_7reg_valEj
12
+ _Z13kernelWrapperPPvjjjj
13
+ __cxa_finalize
14
+ _fini (referenced text)
15
+
16
+
17
+ Call tree stack and functions sizes:
18
+
19
+ stack stack stack call func func function name
20
+ desc level level desc
21
+ ----- ----- ----- ----- ----- ----- --------------------------------------------------------------
22
+ 0 320 0 0 224 1998 _main_init
23
+ 192 320 1 1 1108 1774 _main
24
+ 0 0 2 2 174 174 _ZN3adf11block_writeEPKNS_7reg_valEj
25
+ * * 2 2 * * _Z13kernelWrapperPPvjjjj (_extern)
26
+ 64 128 2 2 324 492 __cxa_finalize
27
+ 64 64 3 3 168 168 _fini
28
+
29
+
30
+ Maximum call level : 3
31
+ Maximum stack level: 3
32
+ Maximum stack size : 320
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.cmic2 ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.cmico ADDED
@@ -0,0 +1 @@
 
 
1
+ +Mdec
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.lst ADDED
@@ -0,0 +1,824 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by darts version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:42:52 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // darts -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -d -h -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -L +Ihex +nanno +u ../Release/0_0 me
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ .text_segment PM 0
9
+ .entry_point
10
+ .label __AIE_ARCH_MODEL_VERSION__21011100__inlined__1__me_basic___main_init_
11
+ .label _main_init
12
+ .function_start
13
+ 0 0x00 0x07 0xb9 0xf5 0x80 0x44 MOVXM sp, #506560
14
+ 6 0x00 0x00 0x04 0x20 0x00 0x44 MOVXM r8, #0
15
+ 12 0x00 0x00 0x08 0x20 0x00 0x44 MOVXM r16, #0
16
+ 18 0x12 0x21 0x07 0x98 EQ r16, r8, r16
17
+ 22 0x80 0x00 0x40 0x40 0x01 0x84 JNZ r16, #128
18
+ .delay_slot
19
+ .swstall delay_slot
20
+ 28 0x00 0x00 NOPX
21
+ .delay_slot
22
+ .swstall delay_slot
23
+ 30 0x00 0x00 NOPX
24
+ .delay_slot
25
+ .swstall delay_slot
26
+ 32 0x00 0x00 NOPX
27
+ .delay_slot
28
+ .swstall delay_slot
29
+ 34 0x00 0x00 NOPX
30
+ .delay_slot
31
+ .swstall delay_slot
32
+ 36 0x00 0x00 NOPX
33
+ 38 0x00 0x2c 0xf7 0x80 0x8b 0x01 0x28 0x50 0x72 0xba NOPA; MOVS p7, p0; MOV r9, r1
34
+ 48 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x3f 0xff 0xff 0x37 0xfe 0x10 0x00 0x00 0xe1 NOPA; NOPB; NOPS; MOVXM p6, #-4; NOPV
35
+ .label TGT_F_main_init_64
36
+ .loop_nesting 1
37
+ 64 0xdf 0x83 0xd5 0x39 0x81 0xd4 LDA p0, [p6], #-4; MOV r10, p6
38
+ 70 0x00 0x00 NOPX
39
+ 72 0x00 0x00 NOPX
40
+ 74 0x00 0x00 NOPX
41
+ 76 0x00 0x00 NOPX
42
+ 78 0x00 0x00 NOPX
43
+ 80 0x00 0x00 NOPX
44
+ .no_stack_arguments
45
+ 82 0x10 0x30 0x00 0x18 JL p0
46
+ .delay_slot
47
+ .swstall delay_slot
48
+ 86 0x00 0x00 NOPX
49
+ .delay_slot
50
+ .swstall delay_slot
51
+ 88 0x00 0x00 NOPX
52
+ .delay_slot
53
+ .swstall delay_slot
54
+ 90 0x00 0x00 NOPX
55
+ .delay_slot
56
+ .swstall delay_slot
57
+ 92 0x00 0x00 NOPX
58
+ .delay_slot
59
+ .swstall delay_slot
60
+ 94 0x00 0x00 NOPX
61
+ .return_address
62
+ 96 0x12 0xa0 0x88 0x98 NE r16, r10, r8
63
+ 100 0x80 0x00 0x20 0x40 0x01 0x84 JNZ r16, #64
64
+ .delay_slot
65
+ .swstall delay_slot
66
+ 106 0x00 0x00 NOPX
67
+ .delay_slot
68
+ .swstall delay_slot
69
+ 108 0x00 0x00 NOPX
70
+ .delay_slot
71
+ .swstall delay_slot
72
+ 110 0x00 0x00 NOPX
73
+ .delay_slot
74
+ .swstall delay_slot
75
+ 112 0x00 0x00 NOPX
76
+ .delay_slot
77
+ .swstall delay_slot
78
+ 114 0x00 0x00 NOPX
79
+ .loop_nesting 0
80
+ 116 0x00 0x2c 0xf0 0x00 0x20 0x9c 0x8b 0x00 0x2a 0x50 0x70 0xf6 NOPA; NOPB; MOVS p0, p7; MOV r1, r9
81
+ .label TGT_F_main_init_128
82
+ .no_stack_arguments
83
+ 128 0x00 0x00 0x70 0x00 0x01 0x04 JL #224
84
+ .delay_slot
85
+ .swstall delay_slot
86
+ 134 0x00 0x00 NOPX
87
+ .delay_slot
88
+ .swstall delay_slot
89
+ 136 0x00 0x00 NOPX
90
+ .delay_slot
91
+ .swstall delay_slot
92
+ 138 0x00 0x00 NOPX
93
+ .delay_slot
94
+ .swstall delay_slot
95
+ 140 0x00 0x00 NOPX
96
+ .delay_slot
97
+ .swstall delay_slot
98
+ 142 0x00 0x00 NOPX
99
+ .return_address
100
+ .no_stack_arguments
101
+ 144 0x00 0x03 0x50 0x00 0x01 0x04 JL #1696
102
+ .delay_slot
103
+ 150 0x18 0x60 0x00 0xb8 MOV p0, #0
104
+ .delay_slot
105
+ .swstall delay_slot
106
+ 154 0x00 0x00 NOPX
107
+ .delay_slot
108
+ .swstall delay_slot
109
+ 156 0x00 0x00 NOPX
110
+ .delay_slot
111
+ .swstall delay_slot
112
+ 158 0x00 0x00 NOPX
113
+ .delay_slot
114
+ .swstall delay_slot
115
+ 160 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
116
+ .return_address
117
+ .swstall chess_separator_scheduler
118
+ 176 0x00 0x00 NOPX
119
+ .swstall chess_separator_scheduler
120
+ 178 0x00 0x00 NOPX
121
+ .swstall chess_separator_scheduler
122
+ 180 0x00 0x00 NOPX
123
+ .swstall chess_separator_scheduler
124
+ 182 0x00 0x00 NOPX
125
+ .swstall chess_separator_scheduler
126
+ 184 0x00 0x00 NOPX
127
+ .swstall chess_separator_scheduler
128
+ 186 0x00 0x00 NOPX
129
+ 188 0x10 0x08 0x00 0x18 DONE
130
+ .swstall chess_separator_scheduler
131
+ 192 0x00 0x00 NOPX
132
+ .swstall chess_separator_scheduler
133
+ 194 0x00 0x00 NOPX
134
+ .swstall chess_separator_scheduler
135
+ 196 0x00 0x00 NOPX
136
+ .swstall chess_separator_scheduler
137
+ 198 0x00 0x00 NOPX
138
+ .swstall chess_separator_scheduler
139
+ 200 0x00 0x00 NOPX
140
+ .swstall chess_separator_scheduler
141
+ 202 0x00 0x00 NOPX
142
+ .swstall for_chess_exit
143
+ .exit
144
+ 204 0x00 0x01 0x67 0x98 NOPA
145
+ .label TGT_F_main_init_208
146
+ .loop_nesting 1
147
+ 208 0x00 0x00 0x68 0x00 0x00 0x84 J #208
148
+ .delay_slot
149
+ .swstall delay_slot
150
+ 214 0x00 0x00 NOPX
151
+ .delay_slot
152
+ .swstall delay_slot
153
+ 216 0x00 0x00 NOPX
154
+ .delay_slot
155
+ .swstall delay_slot
156
+ 218 0x00 0x00 NOPX
157
+ .delay_slot
158
+ .swstall delay_slot
159
+ 220 0x00 0x00 NOPX
160
+ .delay_slot
161
+ .swstall delay_slot
162
+ 222 0x00 0x00 NOPX
163
+ .label _main_init__end
164
+ .label _main___func_begin0
165
+ .label _main
166
+ .function_start
167
+ 224 0xec 0x00 0x80 0x00 0x06 0x00 0x00 0x00 0x70 0xba MOVA m0, #-160; PADDXM [sp], #192
168
+ 234 0xe9 0x04 0x80 0x00 0xc0 0x48 0xb2 0xf0 0x78 0xba MOVA m1, #-184; MOVX r12, #2; MOV p1, sp
169
+ 244 0x00 0x19 0x02 0x17 0x20 0x00 0x4f 0x86 0x0e 0x02 0xd0 0x91 0x60 0x7e MOVA r25, #0; PADDB [p1], m0; MOVS p6, p1; MOVXM p0, #651488
170
+ 258 0x00 0x2c 0xfc 0x57 0x20 0x20 0x05 0x61 0x00 0x0b 0xff 0x93 0xb0 0x7e NOPA; PADDB [p6], m1; ST p1, [sp, #-4]; MOVX r16, #1; MOV r24, #0
171
+ .label TGT_F_main_48
172
+ .loop_nesting 1
173
+ 272 0x08 0x4f 0x3e 0x98 ST.TM r25, [p0], #16
174
+ 276 0x08 0xcf 0x3e 0x98 ST.TM r25, [p0], #-16
175
+ 280 0x00 0x00 NOPX
176
+ 282 0x00 0x00 NOPX
177
+ 284 0x00 0x00 NOPX
178
+ 286 0x17 0xc3 0x08 0x18 ACQ #62, r16
179
+ 290 0xf4 0x9c 0x80 0x00 0x01 0xef 0xb1 0x40 0x10 0xba MOVA m7, #-92; MOVXM p7, #504448
180
+ 300 0x00 0x00 NOPX
181
+ 302 0x00 0x00 NOPX
182
+ 304 0x0f 0xf8 0x1d 0x98 ST p0, [sp, #-8]
183
+ 308 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
184
+ 312 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
185
+ 316 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
186
+ 320 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
187
+ 324 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
188
+ 328 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
189
+ 332 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
190
+ 336 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
191
+ 340 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
192
+ 344 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
193
+ 348 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
194
+ 352 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
195
+ 356 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
196
+ 360 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
197
+ 364 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
198
+ 368 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
199
+ 372 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
200
+ 376 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
201
+ 380 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
202
+ 384 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
203
+ 388 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
204
+ 392 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
205
+ 396 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
206
+ 400 0x09 0xeb 0x31 0x98 ST r25, [p1], m7
207
+ 404 0x00 0x2c 0xf0 0x00 0x20 0x00 0x00 0x03 0xfa 0x4e 0xc1 0x36 NOPA; NOPB; ST p1, [sp, #-12]; NOPX
208
+ .label TGT_F_main_192
209
+ .loop_nesting 2
210
+ 416 0xe3 0xc6 0xd0 0x00 0x00 0x00 0x78 0xf8 0x10 0xba LDA r17, [p7], #4; MOVXM ls, #496
211
+ 426 0xe3 0xaa 0xd0 0x00 0x00 0x01 0xb9 0x28 0x10 0xba LDA r10, [p7], #4; MOVXM le, #592
212
+ 436 0x07 0x1d 0x36 0x98 LDA r9, [p7], #4
213
+ 440 0x07 0x1d 0x16 0x98 LDA r8, [p7], #4
214
+ 444 0x07 0x2e 0x76 0x98 LDA r19, [p7], #8
215
+ 448 0xff 0x93 0x24 0xdd 0x81 0xd4 LDA p1, [sp, #-4]; MOV p2, p7
216
+ 454 0x00 0x00 NOPX
217
+ 456 0x04 0x00 0x27 0x31 0x39 0xe4 MOVX r16, #0; MOV el7, r24
218
+ 462 0x1c 0x90 0x9c 0xf8 MOV el9, r16
219
+ 466 0x1d 0x14 0xa9 0x58 ADD.NC r20, r9, r10
220
+ 470 0x00 0x2c 0xf0 0x98 0x8b 0x02 0x8d 0x10 0xa2 0xba NOPA; MOVS p0, p6; ADD.NC r20, r20, r8
221
+ 480 0x00 0xd2 0x00 0x00 0x20 0x01 0x5b 0x01 0x30 0x0a 0xbc 0xe8 0xa8 0x00 0x00 0xe1 MOVA r18, #6; NOPB; NOPS; MOVX r19, #0; ADD.NC lc, r19, r20; NOPV
222
+ .label ZLS_F_main_272
223
+ .loop_nesting 3
224
+ .begin_of_loop
225
+ 496 0x43 0xe4 0xd1 0x1d 0xe9 0x82 0x6c 0xc0 0x42 0xba LDA dn6, [p2], #4; ST el7, [p1], #4; ADD.NC r19, r19, #1
226
+ 506 0x43 0xd2 0xd9 0x32 0x04 0x14 LDA r20, [p2], #4; ADD.NC r18, r18, #4
227
+ 512 0x02 0x1f 0xe6 0x98 LDA dc7, [p2], #4
228
+ 516 0x02 0x1e 0xee 0x98 LDA el11, [p2], #4
229
+ 520 0x00 0x00 NOPX
230
+ 522 0x00 0x00 NOPX
231
+ 524 0x00 0x00 NOPX
232
+ 526 0x09 0x1f 0x21 0x98 ST dn6, [p1], #4
233
+ 530 0x00 0x00 NOPX
234
+ 532 0x09 0x1f 0xe1 0x98 ST dc7, [p1], #4
235
+ 536 0x23 0xcd 0x30 0x00 0x01 0xa5 0x70 0x02 ST el9, [p1], #4; NOPM
236
+ 544 0x00 0x2c 0xf0 0x00 0x21 0x1e 0x91 0x80 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; ST r20, [p1], #4; NOPX; NOPM; NOPV
237
+ 560 0x00 0x2c 0xf0 0x00 0x21 0xbe 0xe9 0x80 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; ST el11, [p1], #-20; NOPX; NOPM; NOPV
238
+ 576 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x02 0xa9 0x60 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; MOV r21, p1; NOPV
239
+ .label ZLE_F_main_368
240
+ .end_of_loop
241
+ 592 0x00 0x2c 0xf0 0x00 0x20 0x1e 0xb1 0x80 0x00 0x00 0xb5 0x46 0x08 0x00 0x00 0xe1 NOPA; NOPB; ST r21, [p0], #4; NOPX; ADD.NC p1, r21, #24; NOPV
242
+ .loop_nesting 2
243
+ 608 0x00 0x0b 0x00 0x27 0x46 0x6e 0x6f 0x60 0x78 0xba MOVA r11, #0; LSHL r20, r19, r12; MOV r19, p7
244
+ 618 0x9b 0x74 0x69 0xb2 0x01 0x24 ADD r13, r19, #-24; ADD.NC r19, r18, #1
245
+ 624 0xf1 0xa1 0x60 0x25 0x26 0x6f 0x45 0x10 0x79 0x3a MOVS p7, r13; LSHL r18, r18, r12; MOV dj6, r20
246
+ 634 0x9c 0x99 0xbb 0x12 0x41 0xe4 LSHL r18, r19, r12; MOV dj5, r18
247
+ 640 0xf4 0x4a 0xde 0x8d 0x92 0x94 LDA r18, [p7, dj5]; ADD.NC dn7, r13, r18
248
+ 646 0xd8 0x74 0x30 0x02 0x2c 0x7f 0xc0 0x02 ST dn7, [p6, dj6]; ADD.NC r17, r17, #-1
249
+ 654 0x00 0x00 NOPX
250
+ 656 0x00 0x00 NOPX
251
+ 658 0x00 0x00 NOPX
252
+ 660 0x00 0x00 NOPX
253
+ 662 0x00 0x00 NOPX
254
+ 664 0x00 0x2b 0x60 0x02 0x4c 0xe4 0xa0 0x02 NOPS; ADD.NC r18, r19, r18
255
+ .label TGT_F_main_448
256
+ .loop_nesting 3
257
+ 672 0x14 0xa1 0x01 0x98 SUB r16, r18, r16
258
+ 676 0x14 0x1e 0xcd 0x98 LSHL r15, r16, r12
259
+ 680 0xf1 0xa1 0x60 0x03 0xc3 0xd0 0x70 0x02 MOVS p7, r13; MOV dj7, r15
260
+ 688 0x07 0xe0 0x16 0x98 LDA r0, [p7, dj7]
261
+ .no_stack_arguments
262
+ 692 0x00 0x02 0xa0 0x00 0x01 0x04 JL #1344
263
+ .delay_slot
264
+ 698 0x1b 0x98 0x00 0x98 ADD.NC r14, r16, #1
265
+ .delay_slot
266
+ .swstall delay_slot
267
+ 702 0x00 0x00 NOPX
268
+ .delay_slot
269
+ 704 0x13 0xa0 0xcd 0x98 LSHL r16, r14, r12
270
+ .delay_slot
271
+ 708 0x18 0x66 0xc1 0x58 ADD.NC p0, r13, r16
272
+ .delay_slot
273
+ 712 0xfe 0x46 0xb0 0x00 0x01 0xa5 0x70 0x02 ST r17, [sp, #-16]; NOPM
274
+ .return_address
275
+ 720 0x04 0x40 0xa1 0x0f 0x41 0xe4 MOVX r17, #1; MOV dj0, r15
276
+ 726 0xe0 0x4a 0xd0 0x00 0x00 0x1e 0x8b 0xe0 0x10 0xba LDA r18, [p7, dj0]; MOVXM r20, #30656
277
+ 736 0x00 0x00 0x7a 0xaf 0x84 0x44 MOVXM r21, #30658
278
+ 742 0x00 0x08 0x00 0xc0 0x00 0x44 MOVXM p0, #524288
279
+ 748 0x00 0x70 0x08 0x20 0x06 0x44 MOVXM r16, #7340035
280
+ 754 0x00 0x00 NOPX
281
+ 756 0x00 0x00 NOPX
282
+ 758 0x00 0x00 NOPX
283
+ 760 0x14 0xa3 0x1d 0x98 LSHL r17, r18, r17
284
+ 764 0x1c 0x98 0xb9 0x58 ADD.NC r18, r17, r14
285
+ 768 0x14 0xa6 0xcd 0x98 LSHL r19, r18, r12
286
+ 772 0x18 0x89 0xa0 0xf8 MOV dj0, r19
287
+ 776 0x07 0x02 0xd6 0x98 LDA r22, [p7, dj0]
288
+ 780 0x00 0x00 NOPX
289
+ 782 0x00 0x00 NOPX
290
+ 784 0x00 0x00 NOPX
291
+ 786 0x00 0x00 NOPX
292
+ 788 0x00 0x00 NOPX
293
+ 790 0x00 0x00 NOPX
294
+ 792 0x15 0xb6 0xcc 0x98 LTU r27, r22, r12
295
+ 796 0x15 0x69 0x42 0x18 SEL.EQZ r20, r21, r20, r27
296
+ 800 0x1d 0x1b 0x51 0x58 ADD.NC r20, r22, r20
297
+ 804 0x15 0x28 0xcd 0x98 LSHL r20, r20, r12
298
+ 808 0x00 0x2b 0x60 0x00 0x45 0x10 0x70 0x02 NOPS; MOV dj0, r20
299
+ .label TGT_F_main_592
300
+ .loop_nesting 4
301
+ 816 0x00 0x02 0x93 0x98 LDA.TM r20, [p0, dj0]
302
+ 820 0x00 0x00 NOPX
303
+ 822 0x00 0x00 NOPX
304
+ 824 0x00 0x00 NOPX
305
+ 826 0x00 0x00 NOPX
306
+ 828 0x00 0x00 NOPX
307
+ 830 0x00 0x00 NOPX
308
+ 832 0x15 0x2b 0x04 0x98 AND r21, r20, r16
309
+ 836 0xa8 0x01 0x98 0x40 0x01 0x84 JNZ r21, #816
310
+ .delay_slot
311
+ .swstall delay_slot
312
+ 842 0x00 0x00 NOPX
313
+ .delay_slot
314
+ .swstall delay_slot
315
+ 844 0x00 0x00 NOPX
316
+ .delay_slot
317
+ .swstall delay_slot
318
+ 846 0x00 0x00 NOPX
319
+ .delay_slot
320
+ .swstall delay_slot
321
+ 848 0x00 0x00 NOPX
322
+ .delay_slot
323
+ .swstall delay_slot
324
+ 850 0x00 0x00 NOPX
325
+ .loop_nesting 3
326
+ 852 0x1b 0xd9 0x00 0x98 ADD.NC r15, r18, #1
327
+ 856 0x13 0xdc 0xcd 0x98 LSHL r14, r15, r12
328
+ 860 0x19 0x87 0x20 0xf8 MOV dj1, r14
329
+ 864 0x07 0x20 0x16 0x98 LDA r0, [p7, dj1]
330
+ .no_stack_arguments
331
+ 868 0x00 0x02 0xa0 0x00 0x01 0x04 JL #1344
332
+ .delay_slot
333
+ .swstall delay_slot
334
+ 874 0x00 0x00 NOPX
335
+ .delay_slot
336
+ .swstall delay_slot
337
+ 876 0x00 0x00 NOPX
338
+ .delay_slot
339
+ 878 0x13 0x5a 0x23 0x18 ADD r13, r13, #8
340
+ .delay_slot
341
+ 882 0x18 0x69 0xb5 0x58 ADD.NC p0, r19, r13
342
+ .delay_slot
343
+ 886 0x00 0x2c 0xf7 0xea 0x35 0x80 0x00 0x00 0x00 0x7a NOPA; ST r17, [sp, #-24]; NOPX
344
+ .return_address
345
+ 896 0x00 0x0e 0x00 0x01 0x00 0x28 0x43 0x90 0x78 0xba MOVA r14, #0; MOVX r16, #1; MOV dj0, r14
346
+ 906 0xe0 0x4a 0xd8 0xad 0xfc 0x14 LDA r18, [p7, dj0]; ADD.NC r17, r13, #-4
347
+ 912 0x0f 0xee 0x35 0x98 ST r17, [sp, #-20]
348
+ 916 0x00 0x00 NOPX
349
+ 918 0x00 0x00 NOPX
350
+ 920 0x00 0x00 NOPX
351
+ 922 0x00 0x00 NOPX
352
+ 924 0x00 0x00 NOPX
353
+ 926 0x14 0xa5 0x0d 0x98 LSHL r18, r18, r16
354
+ 930 0x14 0xa1 0x05 0x98 OR r16, r18, r16
355
+ 934 0x1c 0x97 0xc1 0x58 ADD.NC r18, r15, r16
356
+ 938 0xfc 0x42 0xb0 0x24 0xc6 0x6c 0x37 0x60 0x79 0x3a ST r16, [sp, #-32]; LSHL r12, r18, r12; MOV p0, p7
357
+ 948 0xfc 0xca 0xb0 0x23 0x06 0x04 0x43 0x10 0x79 0x3a ST r18, [sp, #-28]; ADD r16, r17, r12; MOV dj0, r12
358
+ 958 0x00 0x42 0xd7 0xde 0x15 0x80 0x00 0x03 0xb1 0xf0 0x10 0x76 LDA r16, [p0, dj0]; ST r16, [sp, #-36]; MOVXM p7, #992
359
+ 970 0x00 0x00 NOPX
360
+ 972 0x00 0x00 NOPX
361
+ 974 0x00 0x00 NOPX
362
+ 976 0x00 0x00 NOPX
363
+ 978 0x00 0x00 NOPX
364
+ 980 0x00 0x00 NOPX
365
+ 982 0x00 0x2c 0xf0 0x00 0x10 0x01 0xec 0x3f 0xce 0xba NOPA; NOPB; ADD.NC r15, r16, #-1
366
+ .label TGT_F_main_768
367
+ .loop_nesting 4
368
+ 992 0x12 0xe0 0xe5 0x98 OR r16, r11, r14
369
+ 996 0x80 0x02 0x28 0x40 0x01 0x84 JNZ r16, #1104
370
+ .delay_slot
371
+ .swstall delay_slot
372
+ 1002 0x00 0x00 NOPX
373
+ .delay_slot
374
+ .swstall delay_slot
375
+ 1004 0x00 0x00 NOPX
376
+ .delay_slot
377
+ .swstall delay_slot
378
+ 1006 0x00 0x00 NOPX
379
+ .delay_slot
380
+ .swstall delay_slot
381
+ 1008 0x00 0x00 NOPX
382
+ .delay_slot
383
+ 1010 0x11 0xa1 0x60 0x00 0xc3 0x10 0x70 0x02 MOVS p0, r13; MOV dj1, r12
384
+ 1018 0x07 0xdc 0x99 0x18 LDA p1, [sp, #-36]
385
+ 1022 0x00 0x00 NOPX
386
+ 1024 0x00 0x00 NOPX
387
+ 1026 0x00 0x00 NOPX
388
+ 1028 0x00 0x00 NOPX
389
+ 1030 0x00 0x00 NOPX
390
+ 1032 0x00 0x00 NOPX
391
+ 1034 0x01 0x06 0x16 0x98 LDA r16, [p1]
392
+ 1038 0x00 0x00 NOPX
393
+ 1040 0x00 0x00 NOPX
394
+ 1042 0x00 0x00 NOPX
395
+ 1044 0x00 0x00 NOPX
396
+ 1046 0x00 0x00 NOPX
397
+ 1048 0x00 0x00 NOPX
398
+ 1050 0x80 0x02 0x28 0x00 0x01 0x84 JZ r16, #1104
399
+ .delay_slot
400
+ .swstall delay_slot
401
+ 1056 0x00 0x00 NOPX
402
+ .delay_slot
403
+ .swstall delay_slot
404
+ 1058 0x00 0x00 NOPX
405
+ .delay_slot
406
+ .swstall delay_slot
407
+ 1060 0x00 0x00 NOPX
408
+ .delay_slot
409
+ .swstall delay_slot
410
+ 1062 0x00 0x00 NOPX
411
+ .delay_slot
412
+ .swstall delay_slot
413
+ 1064 0x00 0x00 NOPX
414
+ .swstall chess_separator_scheduler
415
+ 1066 0x00 0x00 NOPX
416
+ .swstall chess_separator_scheduler
417
+ 1068 0x00 0x00 NOPX
418
+ .swstall chess_separator_scheduler
419
+ 1070 0x00 0x00 NOPX
420
+ .swstall chess_separator_scheduler
421
+ 1072 0x00 0x00 NOPX
422
+ .swstall chess_separator_scheduler
423
+ 1074 0x00 0x00 NOPX
424
+ .swstall chess_separator_scheduler
425
+ 1076 0x00 0x00 NOPX
426
+ 1078 0x10 0x08 0x00 0x18 DONE
427
+ .swstall chess_separator_scheduler
428
+ 1082 0x00 0x00 NOPX
429
+ .swstall chess_separator_scheduler
430
+ 1084 0x00 0x00 NOPX
431
+ .swstall chess_separator_scheduler
432
+ 1086 0x00 0x00 NOPX
433
+ .swstall chess_separator_scheduler
434
+ 1088 0x00 0x00 NOPX
435
+ .swstall chess_separator_scheduler
436
+ 1090 0x00 0x00 NOPX
437
+ .swstall chess_separator_scheduler
438
+ 1092 0x00 0x2c 0xf0 0x00 0x20 0x00 0x00 0x00 0x00 0xad 0x81 0x36 NOPA; NOPB; NOPS; NOPX
439
+ .label TGT_F_main_880
440
+ 1104 0x04 0x02 0xd0 0xd9 0x81 0xd4 LDA r0, [p0, dj1]; MOV p0, p6
441
+ .no_stack_arguments
442
+ 1110 0x00 0x04 0x98 0x00 0x01 0x04 JL #2352
443
+ .delay_slot
444
+ .swstall delay_slot
445
+ 1116 0x00 0x00 NOPX
446
+ .delay_slot
447
+ .swstall delay_slot
448
+ 1118 0x00 0x00 NOPX
449
+ .delay_slot
450
+ 1120 0x18 0x55 0x20 0xf8 MOV r1, r10
451
+ .delay_slot
452
+ 1124 0x18 0x94 0xa0 0xf8 MOV r2, r9
453
+ .delay_slot
454
+ 1128 0x00 0x2b 0x60 0x00 0x6a 0x10 0x70 0x02 NOPS; MOV r3, r8
455
+ .return_address
456
+ 1136 0x13 0xdf 0xe0 0x18 JNZD r15, r15, p7
457
+ .delay_slot
458
+ .swstall delay_slot
459
+ 1140 0x00 0x00 NOPX
460
+ .delay_slot
461
+ .swstall delay_slot
462
+ 1142 0x00 0x00 NOPX
463
+ .delay_slot
464
+ .swstall delay_slot
465
+ 1144 0x00 0x00 NOPX
466
+ .delay_slot
467
+ .swstall delay_slot
468
+ 1146 0x00 0x00 NOPX
469
+ .delay_slot
470
+ 1148 0x1b 0x97 0x00 0x98 ADD.NC r14, r14, #1
471
+ .loop_nesting 3
472
+ 1152 0xfe 0x46 0x20 0x00 0xc0 0x49 0x6a 0xc0 0x48 0xba LDA r17, [sp, #-16]; MOVX r12, #2; ADD.NC r11, r11, #1
473
+ 1162 0xfd 0xca 0x20 0x00 0x00 0x03 0xb1 0x50 0x10 0xba LDA r18, [sp, #-20]; MOVXM p7, #672
474
+ 1172 0x07 0xe6 0x71 0x18 LDA r19, [sp, #-28]
475
+ 1176 0x07 0xe2 0x11 0x18 LDA r16, [sp, #-32]
476
+ 1180 0x07 0xea 0x91 0x18 LDA r20, [sp, #-24]
477
+ 1184 0x00 0x00 NOPX
478
+ 1186 0x00 0x00 NOPX
479
+ 1188 0x14 0x63 0xe0 0x18 JNZD r17, r17, p7
480
+ .delay_slot
481
+ 1192 0x14 0x9b 0xf3 0x18 ADD r13, r18, #-4
482
+ .delay_slot
483
+ 1196 0x1c 0x99 0x81 0x98 ADD.NC r18, r19, #3
484
+ .delay_slot
485
+ .swstall delay_slot
486
+ 1200 0x00 0x00 NOPX
487
+ .delay_slot
488
+ 1202 0x1c 0x1a 0x41 0x58 ADD.NC r16, r20, r16
489
+ .delay_slot
490
+ 1206 0x1c 0x18 0x02 0x98 ADD.NC r16, r16, #5
491
+ .loop_nesting 2
492
+ 1210 0x07 0xd2 0x07 0x8d 0x0b 0x25 0x06 0x6e 0x88 0x3f 0x58 0x76 MOVA r18, #62; MOVS p7, r13; LSHL r16, r18, r12; MOV r20, #63
493
+ 1222 0x00 0x30 0x00 0x3f 0x67 0xe8 0x44 0x10 0x78 0xba MOVA r16, #1; MOVX r22, #-1; MOV dj0, r16
494
+ 1232 0xe0 0x46 0xd0 0x00 0x01 0xee 0x69 0x40 0x10 0xba LDA r17, [p7, dj0]; MOVXM r19, #504448
495
+ 1242 0x9e 0xda 0xfc 0x20 0x01 0x64 EQ r27, r19, r13; MOV r24, #0
496
+ 1248 0x15 0x2f 0x22 0x18 SEL.EQZ r23, r20, r18, r27
497
+ 1252 0x00 0x07 0xba 0xad 0x00 0x44 MOVXM r21, #505472
498
+ 1258 0x00 0x00 NOPX
499
+ 1260 0x00 0x00 NOPX
500
+ 1262 0x00 0x00 NOPX
501
+ 1264 0x14 0x74 0xd0 0x18 EQZ r26, r17
502
+ 1268 0x15 0xd1 0x68 0x18 REL r23, r22
503
+ 1272 0x14 0xa5 0x42 0x18 SEL.EQZ r18, r18, r20, r27
504
+ 1276 0x00 0x00 NOPX
505
+ 1278 0x00 0x00 NOPX
506
+ 1280 0x14 0x97 0x08 0x18 ACQ.COND r18, r16, r26
507
+ 1284 0x88 0x00 0xd0 0x00 0x01 0x84 JZ r17, #416
508
+ .delay_slot
509
+ 1290 0x14 0xe1 0x52 0x18 SEL.EQZ r16, r19, r21, r27
510
+ .delay_slot
511
+ 1294 0x1f 0x68 0x20 0xf8 MOV p7, r16
512
+ .delay_slot
513
+ .swstall delay_slot
514
+ 1298 0x00 0x00 NOPX
515
+ .delay_slot
516
+ .swstall delay_slot
517
+ 1300 0x00 0x00 NOPX
518
+ .delay_slot
519
+ .swstall delay_slot
520
+ 1302 0x00 0x00 NOPX
521
+ .loop_nesting 1
522
+ 1304 0xff 0x03 0x20 0x01 0x90 0x0a 0x08 0x01 0x58 0xba LDA p0, [sp, #-8]; MOVX r25, #0; MOV r16, #1
523
+ 1314 0x00 0x00 0x88 0x00 0x00 0x84 J #272
524
+ .delay_slot
525
+ .swstall delay_slot
526
+ 1320 0x00 0x00 NOPX
527
+ .delay_slot
528
+ .swstall delay_slot
529
+ 1322 0x00 0x00 NOPX
530
+ .delay_slot
531
+ .swstall delay_slot
532
+ 1324 0x00 0x00 NOPX
533
+ .delay_slot
534
+ .swstall delay_slot
535
+ 1326 0x00 0x00 NOPX
536
+ .delay_slot
537
+ 1328 0x07 0xf4 0x99 0x18 LDA p1, [sp, #-12]
538
+ .label _main__end
539
+ .label _main___func_end0
540
+
541
+ .text_segment PM 1344
542
+ .label __ZN3adf11block_writeEPKNS_7reg_valEj___func_begin0
543
+ .label _ZN3adf11block_writeEPKNS_7reg_valEj
544
+ .function_start
545
+ 1344 0x00 0x02 0xf0 0x00 0x01 0x84 JZ r0, #1504
546
+ .delay_slot
547
+ .swstall delay_slot
548
+ 1350 0x00 0x00 NOPX
549
+ .delay_slot
550
+ .swstall delay_slot
551
+ 1352 0x00 0x00 NOPX
552
+ .delay_slot
553
+ .swstall delay_slot
554
+ 1354 0x00 0x00 NOPX
555
+ .delay_slot
556
+ .swstall delay_slot
557
+ 1356 0x00 0x00 NOPX
558
+ .delay_slot
559
+ .swstall delay_slot
560
+ 1358 0x00 0x00 NOPX
561
+ 1360 0xf8 0x1e 0x2a 0xe0 0x41 0xe4 MOVX r0, #-4; MOV lc, r0
562
+ 1366 0x00 0x00 0x01 0xea 0xe0 0x44 MOVXM ls, #1392
563
+ 1372 0x00 0x00 0x06 0xeb 0xa0 0x44 MOVXM le, #1488
564
+ 1378 0x00 0x2c 0xf0 0x00 0x20 0x00 0x40 0x16 0x00 0x02 0x00 0x2b 0x60 0x7e NOPA; NOPB; NOPS; MOVXM p1, #524288
565
+ .label ZLS_F_ZN3adf11block_writeEPKNS_7reg_valEj_48
566
+ .loop_nesting 1
567
+ .begin_of_loop
568
+ 1392 0x00 0x1c 0x56 0x98 LDA r2, [p0], #4
569
+ 1396 0x00 0x1c 0x36 0x98 LDA r1, [p0], #4
570
+ 1400 0x00 0x00 NOPX
571
+ 1402 0x00 0x2c 0xf0 0x00 0x20 0x3c NOPA; NOPB
572
+ 1408 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
573
+ 1424 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
574
+ 1440 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
575
+ 1456 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x04 0x30 0x24 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; AND r3, r2, r0; NOPM; NOPV
576
+ 1472 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x40 0xd0 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; MOV dj0, r3; NOPV
577
+ .label ZLE_F_ZN3adf11block_writeEPKNS_7reg_valEj_144
578
+ .end_of_loop
579
+ 1488 0x00 0x2c 0xf0 0x00 0x21 0x00 0x3e 0x80 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; ST.TM r1, [p1, dj0]; NOPX; NOPM; NOPV
580
+ .label TGT_F_ZN3adf11block_writeEPKNS_7reg_valEj_160
581
+ .loop_nesting 0
582
+ 1504 0x10 0x28 0x00 0x18 RET lr
583
+ .delay_slot
584
+ .swstall delay_slot
585
+ 1508 0x00 0x00 NOPX
586
+ .delay_slot
587
+ .swstall delay_slot
588
+ 1510 0x00 0x00 NOPX
589
+ .delay_slot
590
+ .swstall delay_slot
591
+ 1512 0x00 0x00 NOPX
592
+ .delay_slot
593
+ .swstall delay_slot
594
+ 1514 0x00 0x00 NOPX
595
+ .delay_slot
596
+ .swstall delay_slot
597
+ 1516 0x00 0x00 NOPX
598
+ .label _ZN3adf11block_writeEPKNS_7reg_valEj__end
599
+ .label __ZN3adf11block_writeEPKNS_7reg_valEj___func_end0
600
+
601
+ .text_segment PM 1520
602
+ .label _fini
603
+ .function_start
604
+ 1520 0x00 0x08 0x00 0x00 0x01 0xc4 PADDXM [sp], #64
605
+ 1526 0xfe 0xba 0xb0 0x00 0x00 0x02 0x08 0x00 0x11 0x3a ST r14, [sp, #-12]; MOVXM r16, #0
606
+ 1536 0xff 0x73 0xb0 0x00 0x00 0x01 0xc8 0x00 0x11 0x3a ST p7, [sp, #-8]; MOVXM r14, #0
607
+ 1546 0x74 0x20 0xfe 0xd0 0x41 0xe4 EQ r16, r14, r16; MOV p7, r16
608
+ 1552 0x80 0x03 0x38 0x40 0x01 0x84 JNZ r16, #1648
609
+ .delay_slot
610
+ 1558 0x0f 0xfd 0xf5 0x98 ST r15, [sp, #-4]
611
+ .delay_slot
612
+ 1562 0x0f 0xf0 0x3d 0x98 ST lr, [sp, #-16]
613
+ .delay_slot
614
+ .swstall delay_slot
615
+ 1566 0x00 0x00 NOPX
616
+ .delay_slot
617
+ .swstall delay_slot
618
+ 1568 0x00 0x00 NOPX
619
+ .delay_slot
620
+ .swstall delay_slot
621
+ 1570 0x00 0x2c 0xf0 0x02 0xb6 0x00 0x00 0x40 0x1a 0x57 0x00 0x00 0x1c 0x2e NOPA; NOPS; NOPM; NOPV
622
+ .label TGT_F_fini_64
623
+ .loop_nesting 1
624
+ 1584 0x07 0x1c 0x1e 0x98 LDA p0, [p7], #4
625
+ 1588 0x00 0x00 NOPX
626
+ 1590 0x00 0x00 NOPX
627
+ 1592 0x00 0x00 NOPX
628
+ 1594 0x00 0x00 NOPX
629
+ 1596 0x00 0x00 NOPX
630
+ 1598 0x00 0x00 NOPX
631
+ .no_stack_arguments
632
+ 1600 0x10 0x30 0x00 0x18 JL p0
633
+ .delay_slot
634
+ 1604 0x1b 0xde 0xc0 0xf8 MOV r15, p7
635
+ .delay_slot
636
+ .swstall delay_slot
637
+ 1608 0x00 0x00 NOPX
638
+ .delay_slot
639
+ .swstall delay_slot
640
+ 1610 0x00 0x00 NOPX
641
+ .delay_slot
642
+ .swstall delay_slot
643
+ 1612 0x00 0x00 NOPX
644
+ .delay_slot
645
+ .swstall delay_slot
646
+ 1614 0x00 0x00 NOPX
647
+ .return_address
648
+ 1616 0x13 0xe0 0xe8 0x98 NE r16, r15, r14
649
+ 1620 0x80 0x03 0x18 0x40 0x01 0x84 JNZ r16, #1584
650
+ .delay_slot
651
+ .swstall delay_slot
652
+ 1626 0x00 0x00 NOPX
653
+ .delay_slot
654
+ .swstall delay_slot
655
+ 1628 0x00 0x00 NOPX
656
+ .delay_slot
657
+ .swstall delay_slot
658
+ 1630 0x00 0x00 NOPX
659
+ .delay_slot
660
+ .swstall delay_slot
661
+ 1632 0x00 0x00 NOPX
662
+ .delay_slot
663
+ .swstall delay_slot
664
+ 1634 0x00 0x2c 0xf0 0x02 0xb6 0x00 0x00 0x40 0x1a 0x57 0x00 0x00 0x1c 0x2e NOPA; NOPS; NOPM; NOPV
665
+ .label TGT_F_fini_128
666
+ .loop_nesting 0
667
+ 1648 0x07 0xf0 0x39 0x18 LDA lr, [sp, #-16]
668
+ 1652 0x07 0xf5 0xd1 0x18 LDA r14, [sp, #-12]
669
+ 1656 0x07 0xfd 0xf1 0x18 LDA r15, [sp, #-4]
670
+ 1660 0x07 0xfb 0x99 0x18 LDA p7, [sp, #-8]
671
+ 1664 0x00 0x00 NOPX
672
+ 1666 0x00 0x00 NOPX
673
+ 1668 0x00 0x00 NOPX
674
+ 1670 0x10 0x28 0x00 0x18 RET lr
675
+ .delay_slot
676
+ 1674 0xff 0xf8 0x00 0x00 0x01 0xc4 PADDXM [sp], #-64
677
+ .delay_slot
678
+ .swstall delay_slot
679
+ 1680 0x00 0x00 NOPX
680
+ .delay_slot
681
+ .swstall delay_slot
682
+ 1682 0x00 0x00 NOPX
683
+ .delay_slot
684
+ .swstall delay_slot
685
+ 1684 0x00 0x00 NOPX
686
+ .delay_slot
687
+ .swstall delay_slot
688
+ 1686 0x00 0x00 NOPX
689
+ .label _fini__end
690
+
691
+ .text_segment PM 1696
692
+ .label __cxa_finalize
693
+ .function_start
694
+ 1696 0xff 0xa1 0x00 0x00 0x02 0x00 0x00 0x00 0x70 0xba MOVA r1, #-3; PADDXM [sp], #64
695
+ 1706 0xfb 0x07 0xb0 0x00 0x41 0x08 0x6d 0xe0 0x79 0x3a ST lr, [sp, #-40]; MOVX r4, #8; MOV r3, packSign1
696
+ 1716 0xfa 0x8e 0xb0 0x00 0x69 0xe0 0x70 0x02 ST r3, [sp, #-44]; MOV r3, packSign0
697
+ 1724 0xfc 0x0e 0xb0 0x00 0x6e 0x30 0x70 0x02 ST r3, [sp, #-32]; MOV r3, unpackSign1
698
+ 1732 0xfc 0x8e 0xb0 0x00 0x6a 0x30 0x70 0x02 ST r3, [sp, #-28]; MOV r3, unpackSign0
699
+ 1740 0xfd 0x8e 0xb0 0x00 0x01 0xf0 0x32 0x10 0x11 0x3a ST r3, [sp, #-20]; MOVXM p0, #508960
700
+ 1750 0x00 0x8a 0xd1 0xb1 0xc1 0xd4 LDA r2, [p0]; MOV r3, crSRSMode
701
+ 1756 0xfe 0x0e 0xb0 0x00 0x6b 0xb0 0x70 0x02 ST r3, [sp, #-16]; MOV r3, crPackSize
702
+ 1764 0xfe 0x8e 0xb0 0x00 0x6a 0x70 0x70 0x02 ST r3, [sp, #-12]; MOV r3, crSat
703
+ 1772 0x00 0x92 0x30 0x01 0x2c 0x30 0x70 0x02 ST r4, [p0]; MOV r9, upsSign1
704
+ 1780 0xff 0x0e 0xb0 0x01 0x08 0x30 0x70 0x02 ST r3, [sp, #-8]; MOV r8, upsSign0
705
+ 1788 0x1a 0xda 0x60 0xf8 MOV r11, vaddSign1
706
+ 1792 0x1a 0x92 0x60 0xf8 MOV r10, vaddSign0
707
+ 1796 0x10 0xbc 0x66 0x3f 0x81 0xe4 ADD r2, r2, #-8; MOV r12, srsSign1
708
+ 1802 0x13 0x43 0xd0 0xa5 0xc1 0xe4 ASHL r13, r2, r1; MOV r1, crUnpackSize
709
+ 1808 0xff 0x86 0xb0 0x00 0x2f 0xb0 0x70 0x02 ST r1, [sp, #-4]; MOV r1, crRnd
710
+ 1816 0xfd 0x06 0xb0 0x00 0x2e 0x70 0x70 0x02 ST r1, [sp, #-24]; MOV r1, crUPSMode
711
+ 1824 0xfb 0x86 0xb0 0x01 0xcb 0xe0 0x70 0x02 ST r1, [sp, #-36]; MOV r14, srsSign0
712
+ 1832 0x00 0x07 0xc0 0xa8 0x00 0x44 MOVXM r1, #508928
713
+ 1838 0x1e 0x60 0x89 0x58 ADD.NC p6, r1, r2
714
+ 1842 0x00 0x2c 0xf0 0x00 0x20 0x00 0x00 0x76 0x74 0x02 0x00 0x2b 0x60 0x7e NOPA; NOPB; NOPS; MOVXM p7, #1856
715
+ .label TGT_F__cxa_finalize_160
716
+ .loop_nesting 1
717
+ 1856 0x06 0x1c 0x9e 0x98 LDA p1, [p6], #4
718
+ 1860 0x06 0xdc 0x1e 0x98 LDA p0, [p6], #-12
719
+ 1864 0x00 0x00 NOPX
720
+ 1866 0x00 0x00 NOPX
721
+ 1868 0x00 0x00 NOPX
722
+ 1870 0x00 0x00 NOPX
723
+ 1872 0x00 0x00 NOPX
724
+ .no_stack_arguments
725
+ 1874 0x10 0x30 0x40 0x18 JL p1
726
+ .delay_slot
727
+ 1878 0x1b 0xd0 0x20 0xf8 MOV r15, r0
728
+ .delay_slot
729
+ .swstall delay_slot
730
+ 1882 0x00 0x00 NOPX
731
+ .delay_slot
732
+ .swstall delay_slot
733
+ 1884 0x00 0x00 NOPX
734
+ .delay_slot
735
+ .swstall delay_slot
736
+ 1886 0x00 0x00 NOPX
737
+ .delay_slot
738
+ .swstall delay_slot
739
+ 1888 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
740
+ .return_address
741
+ 1904 0x13 0x5b 0xe0 0x18 JNZD r13, r13, p7
742
+ .delay_slot
743
+ 1908 0x18 0x17 0xa0 0xf8 MOV r0, r15
744
+ .delay_slot
745
+ .swstall delay_slot
746
+ 1912 0x00 0x00 NOPX
747
+ .delay_slot
748
+ .swstall delay_slot
749
+ 1914 0x00 0x00 NOPX
750
+ .delay_slot
751
+ .swstall delay_slot
752
+ 1916 0x00 0x00 NOPX
753
+ .delay_slot
754
+ .swstall delay_slot
755
+ 1918 0x00 0x00 NOPX
756
+ .loop_nesting 0
757
+ 1920 0xfb 0x07 0x20 0x13 0xea 0x02 0x9a 0xd0 0x78 0xba LDA lr, [sp, #-40]; MOVX upsSign1, r9; MOV vaddSign1, r11
758
+ 1930 0xfa 0x86 0x20 0x11 0xca 0x00 0x9a 0x90 0x78 0xba LDA r1, [sp, #-44]; MOVX upsSign0, r8; MOV vaddSign0, r10
759
+ 1940 0xfc 0x0a 0x26 0x79 0x80 0x2c LDA r2, [sp, #-32]; MOVX srsSign1, r12
760
+ 1946 0xfc 0x8e 0x27 0x71 0x80 0x2c LDA r3, [sp, #-28]; MOVX srsSign0, r14
761
+ 1952 0x07 0xec 0xf1 0x18 LDA r7, [sp, #-20]
762
+ 1956 0x07 0xf0 0x91 0x18 LDA r4, [sp, #-16]
763
+ 1960 0x07 0xf4 0xb1 0x18 LDA r5, [sp, #-12]
764
+ 1964 0x07 0xf8 0xd1 0x18 LDA r6, [sp, #-8]
765
+ 1968 0xff 0x86 0x20 0xf8 0x80 0x2c LDA r1, [sp, #-4]; MOVX packSign1, r1
766
+ 1974 0xfd 0x0a 0x21 0x70 0x80 0x2c LDA r2, [sp, #-24]; MOVX packSign0, r2
767
+ 1980 0xfb 0x8e 0x21 0xfb 0x80 0x2c LDA r3, [sp, #-36]; MOVX unpackSign1, r3
768
+ 1986 0xff 0xf8 0x00 0x00 0x01 0xc4 PADDXM [sp], #-64
769
+ 1992 0x05 0x00 0x04 0x67 0x41 0xe4 RET lr; MOV unpackSign0, r7
770
+ .delay_slot
771
+ 1998 0x27 0x60 0x07 0x65 0x41 0xe4 MOVX crSRSMode, r4; MOV crPackSize, r5
772
+ .delay_slot
773
+ 2004 0x11 0xbb 0x80 0x18 MOVX crSat, r6
774
+ .delay_slot
775
+ 2008 0x10 0x7c 0x80 0x18 MOVX crUnpackSize, r1
776
+ .delay_slot
777
+ 2012 0x10 0xba 0x80 0x18 MOVX crRnd, r2
778
+ .delay_slot
779
+ 2016 0x10 0xfc 0x00 0x18 MOVX crUPSMode, r3
780
+ .label __cxa_finalize__end
781
+
782
+ .bss_segment DMb 504448 2048
783
+
784
+ .data_segment DMb 508928
785
+ .label _ZL7atexits
786
+ 0xf0
787
+ 0x5
788
+ 0x0
789
+ 0x0
790
+ 0x0
791
+ 0x0
792
+ 0x0
793
+ 0x0
794
+ 0x0
795
+ 0x0
796
+ 0x0
797
+ 0x0
798
+ 0x0
799
+ 0x0
800
+ 0x0
801
+ 0x0
802
+ 0x0
803
+ 0x0
804
+ 0x0
805
+ 0x0
806
+ 0x0
807
+ 0x0
808
+ 0x0
809
+ 0x0
810
+ 0x0
811
+ 0x0
812
+ 0x0
813
+ 0x0
814
+ 0x0
815
+ 0x0
816
+ 0x0
817
+ 0x0
818
+ .label _ZL10atexit_cnt
819
+ 0x8
820
+ 0x0
821
+ 0x0
822
+ 0x0
823
+
824
+ .stack DM_stack 506560 508928
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.map ADDED
@@ -0,0 +1,143 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:42:51 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0 ../Release/0_0.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork848 -pme
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ Memory map for memory 'DM_bankA':
9
+
10
+ Size = 1048576
11
+ Width = 8 bits
12
+ Offset = 0
13
+ Used = 36
14
+
15
+ 0x0007c400..0x0007c41f ( 32 items) : atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a)::_ZL7atexits (Data, Local, .data.DM_bankA.4)
16
+
17
+ Called functions : _fini
18
+
19
+ 0x0007c420..0x0007c423 ( 4 items) : atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a)::_ZL10atexit_cnt (Data, Local, .data.DM_bankA.4)
20
+
21
+ Memory map for memory 'DM_stack':
22
+
23
+ Size = 1048576
24
+ Width = 8 bits
25
+ Offset = 0
26
+ Used = 2368
27
+
28
+ 0x0007bac0..0x0007c3ff ( 2368 items) : Stack
29
+
30
+ Memory map for memory 'DMb':
31
+
32
+ Size = 1048576
33
+ Width = 8 bits
34
+ Offset = 0
35
+ Used = 4452
36
+
37
+ 0x00000000..0x0007b27f ( 504448 items) : Reserved
38
+ 0x0007b280..0x0007b67f ( 1024 items) : ../Release/0_0.o::lcpPing (Data, Global, .bss.DMb.4)
39
+ 0x0007b680..0x0007ba7f ( 1024 items) : ../Release/0_0.o::lcpPong (Data, Global, .bss.DMb.4)
40
+ 0x0007ba80..0x0007babf ( 64 items) : Reserved
41
+ 0x0007bac0..0x0007c3ff ( 2368 items) : Stack
42
+ 0x0007c400..0x0007c41f : Occupied in alias or record memory 'DM_bankA' by symbol '_ZL7atexits'
43
+ 0x0007c420..0x0007c423 : Occupied in alias or record memory 'DM_bankA' by symbol '_ZL10atexit_cnt'
44
+ 0x0007c440..0x000fffff ( 539584 items) : Reserved
45
+
46
+ Memory map for memory 'PM':
47
+
48
+ Size = 1048576
49
+ Width = 8 bits
50
+ Offset = 0
51
+ Used = 1998
52
+
53
+ 0x00000000..0x000000df ( 224 items) : me_basic.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_main_init (Function, Global, .text) (stack frame size = 0)
54
+
55
+ Called functions : _main
56
+ __cxa_finalize
57
+
58
+ Referenced symbols: _sp_start_value_DM_stack
59
+ _ctors_start
60
+ _ctors_end
61
+
62
+ 0x000000e0..0x00000533 ( 1108 items) : ../Release/0_0.o::_main (Function, Global, .text) (stack frame size = 192)
63
+
64
+ Called functions : _ZN3adf11block_writeEPKNS_7reg_valEj
65
+ _Z13kernelWrapperPPvjjjj
66
+
67
+ Referenced symbols: lcpPing
68
+ lcpPong
69
+
70
+ 0x00000540..0x000005ed ( 174 items) : ../Release/0_0.o::_ZN3adf11block_writeEPKNS_7reg_valEj (Function, Weak, .text) (stack frame size = 0)
71
+ 0x000005f0..0x00000697 ( 168 items) : me_basic.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)::_fini (Function, Global, .text) (stack frame size = 64)
72
+
73
+ Referenced symbols: _dtors_start
74
+ _dtors_end
75
+
76
+ 0x000006a0..0x000007e3 ( 324 items) : atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a)::__cxa_finalize (Function, Global, .text) (stack frame size = 64)
77
+
78
+ Referenced symbols: _ZL10atexit_cnt
79
+ _ZL7atexits
80
+
81
+ 0x00000930..0x00003fff ( 14032 items) : Reserved
82
+
83
+ External symbols:
84
+
85
+ _Z13kernelWrapperPPvjjjj = 0x930
86
+ __dso_handle = 0x0
87
+ _ctors_end = 0x0
88
+ _ctors_start = 0x0
89
+ _dtors_end = 0x0
90
+ _dtors_start = 0x0
91
+ _pc_end = 0x7e4
92
+ _pc_start = 0x0
93
+ _sp_end_DM_stack = 0x7c400
94
+ _sp_start_DM_stack = 0x7bac0
95
+
96
+ Section summary for memory 'DM_bankA':
97
+
98
+ .data File
99
+ ---------- ----------
100
+ 36 atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a)
101
+ ---------- ----------
102
+ 36 Total
103
+
104
+ Section summary for memory 'DM_stack':
105
+
106
+ .stack File
107
+ ---------- ----------
108
+ 2368
109
+ ---------- ----------
110
+ 2368 Total
111
+
112
+ Section summary for memory 'DMb':
113
+
114
+ .bss .data File
115
+ ---------- ---------- ----------
116
+ 2048 0 ../Release/0_0.o
117
+ 0 36 atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a) (in DM_bankA)
118
+ ---------- ---------- ----------
119
+ 2048 36 Total
120
+
121
+ Section summary for memory 'PM':
122
+
123
+ .text File
124
+ ---------- ----------
125
+ 1282 ../Release/0_0.o
126
+ 392 me_basic.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
127
+ 324 atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a)
128
+ ---------- ----------
129
+ 1998 Total
130
+
131
+ File summary:
132
+
133
+ atexit.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release/libc.a)
134
+ DM_bankA 36
135
+ PM 324
136
+
137
+ ../Release/0_0.o
138
+ DMb 2048
139
+ PM 1282
140
+
141
+ me_basic.o(/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release/libme.a)
142
+ PM 392
143
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.o ADDED
Binary file (52.1 kB). View file
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.o.lst ADDED
@@ -0,0 +1,471 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by darts version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:42:50 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // darts -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -d -h -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -L +Ihex +nanno ../Release/0_0.o me
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+ .data_segment_name
8
+ .bss local .at 524288 _ZL22TM_Address_Space_Start TM 4
9
+
10
+
11
+ .data_segment_name
12
+ .bss global 4 lcpPing DMb 1024
13
+
14
+ .data_segment_name
15
+ .bss global 4 lcpPong DMb 1024
16
+
17
+ .undef global data _ZN12me_primitive11control_satE
18
+
19
+ .undef global data _ZN12me_primitive11control_rndE
20
+
21
+ .text_segment_name
22
+ .text weak 16 _ZN3adf11block_writeEPKNS_7reg_valEj
23
+ .label __ZN3adf11block_writeEPKNS_7reg_valEj___func_begin0
24
+ .function_start
25
+ 0 0x00 0x00 0x50 0x00 0x01 0x84 JZ r0, #TGT_F_ZN3adf11block_writeEPKNS_7reg_valEj_160
26
+ .delay_slot
27
+ .swstall delay_slot
28
+ 6 0x00 0x00 NOPX
29
+ .delay_slot
30
+ .swstall delay_slot
31
+ 8 0x00 0x00 NOPX
32
+ .delay_slot
33
+ .swstall delay_slot
34
+ 10 0x00 0x00 NOPX
35
+ .delay_slot
36
+ .swstall delay_slot
37
+ 12 0x00 0x00 NOPX
38
+ .delay_slot
39
+ .swstall delay_slot
40
+ 14 0x00 0x00 NOPX
41
+ 16 0xf8 0x1e 0x2a 0xe0 0x41 0xe4 MOVX r0, #-4; MOV lc, r0
42
+ 22 0x00 0x00 0x01 0xe0 0x60 0x44 MOVXM ls, #(ZLS_F_ZN3adf11block_writeEPKNS_7reg_valEj_48 + 0)
43
+ 28 0x00 0x00 0x06 0xe1 0x20 0x44 MOVXM le, #(ZLE_F_ZN3adf11block_writeEPKNS_7reg_valEj_144 + 0)
44
+ 34 0x00 0x2c 0xf0 0x00 0x20 0x00 0x40 0x16 0x00 0x02 0x00 0x2b 0x60 0x7e NOPA; NOPB; NOPS; MOVXM p1, #524288
45
+ .label ZLS_F_ZN3adf11block_writeEPKNS_7reg_valEj_48
46
+ .loop_nesting 1
47
+ .begin_of_loop
48
+ 48 0x00 0x1c 0x56 0x98 LDA r2, [p0], #4
49
+ 52 0x00 0x1c 0x36 0x98 LDA r1, [p0], #4
50
+ 56 0x00 0x00 NOPX
51
+ 58 0x00 0x2c 0xf0 0x00 0x20 0x3c NOPA; NOPB
52
+ 64 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
53
+ 80 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
54
+ 96 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; NOPM; NOPV
55
+ 112 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x04 0x30 0x24 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; AND r3, r2, r0; NOPM; NOPV
56
+ 128 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x00 0x40 0xd0 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; MOV dj0, r3; NOPV
57
+ .label ZLE_F_ZN3adf11block_writeEPKNS_7reg_valEj_144
58
+ .end_of_loop
59
+ 144 0x00 0x2c 0xf0 0x00 0x21 0x00 0x3e 0x80 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; ST.TM r1, [p1, dj0]; NOPX; NOPM; NOPV
60
+ .label TGT_F_ZN3adf11block_writeEPKNS_7reg_valEj_160
61
+ .loop_nesting 0
62
+ 160 0x10 0x28 0x00 0x18 RET lr
63
+ .delay_slot
64
+ .swstall delay_slot
65
+ 164 0x00 0x00 NOPX
66
+ .delay_slot
67
+ .swstall delay_slot
68
+ 166 0x00 0x00 NOPX
69
+ .delay_slot
70
+ .swstall delay_slot
71
+ 168 0x00 0x00 NOPX
72
+ .delay_slot
73
+ .swstall delay_slot
74
+ 170 0x00 0x00 NOPX
75
+ .label _ZN3adf11block_writeEPKNS_7reg_valEj__end last
76
+ .label __ZN3adf11block_writeEPKNS_7reg_valEj___func_end0 last
77
+ .delay_slot
78
+ .swstall delay_slot
79
+ 172 0x00 0x00 NOPX
80
+
81
+ .text_segment_name
82
+ .text global 10 _main
83
+ .label _main___func_begin0
84
+ .function_start
85
+ 0 0xec 0x00 0x80 0x00 0x06 0x00 0x00 0x00 0x70 0xba MOVA m0, #-160; PADDXM [sp], #192
86
+ 10 0xe9 0x04 0x80 0x00 0xc0 0x48 0xb2 0xf0 0x78 0xba MOVA m1, #-184; MOVX r12, #2; MOV p1, sp
87
+ 20 0x00 0x19 0x02 0x17 0x20 0x00 0x4f 0x86 0x0e 0x02 0xd0 0x91 0x60 0x7e MOVA r25, #0; PADDB [p1], m0; MOVS p6, p1; MOVXM p0, #651488
88
+ 34 0x00 0x2c 0xfc 0x57 0x20 0x20 0x05 0x61 0x00 0x0b 0xff 0x93 0xb0 0x7e NOPA; PADDB [p6], m1; ST p1, [sp, #-4]; MOVX r16, #1; MOV r24, #0
89
+ .label TGT_F_main_48
90
+ .loop_nesting 1
91
+ 48 0x08 0x4f 0x3e 0x98 ST.TM r25, [p0], #16
92
+ 52 0x08 0xcf 0x3e 0x98 ST.TM r25, [p0], #-16
93
+ 56 0x00 0x00 NOPX
94
+ 58 0x00 0x00 NOPX
95
+ 60 0x00 0x00 NOPX
96
+ 62 0x17 0xc3 0x08 0x18 ACQ #62, r16
97
+ 66 0xf4 0x9c 0x80 0x00 0x00 0x03 0xb0 0x00 0x10 0xba MOVA m7, #-92; MOVXM p7, #lcpPing
98
+ 76 0x00 0x00 NOPX
99
+ 78 0x00 0x00 NOPX
100
+ 80 0x0f 0xf8 0x1d 0x98 ST p0, [sp, #-8]
101
+ 84 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
102
+ 88 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
103
+ 92 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
104
+ 96 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
105
+ 100 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
106
+ 104 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
107
+ 108 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
108
+ 112 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
109
+ 116 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
110
+ 120 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
111
+ 124 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
112
+ 128 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
113
+ 132 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
114
+ 136 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
115
+ 140 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
116
+ 144 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
117
+ 148 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
118
+ 152 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
119
+ 156 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
120
+ 160 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
121
+ 164 0x09 0x1f 0x11 0x98 ST r24, [p1], #4
122
+ 168 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
123
+ 172 0x09 0x1f 0x31 0x98 ST r25, [p1], #4
124
+ 176 0x09 0xeb 0x31 0x98 ST r25, [p1], m7
125
+ 180 0x00 0x2c 0xf0 0x00 0x20 0x00 0x00 0x03 0xfa 0x4e 0xc1 0x36 NOPA; NOPB; ST p1, [sp, #-12]; NOPX
126
+ .label TGT_F_main_192
127
+ .loop_nesting 2
128
+ 192 0xe3 0xc6 0xd0 0x00 0x00 0x00 0x78 0x88 0x10 0xba LDA r17, [p7], #4; MOVXM ls, #(ZLS_F_main_272 + 0)
129
+ 202 0xe3 0xaa 0xd0 0x00 0x00 0x01 0xb8 0xb8 0x10 0xba LDA r10, [p7], #4; MOVXM le, #(ZLE_F_main_368 + 0)
130
+ 212 0x07 0x1d 0x36 0x98 LDA r9, [p7], #4
131
+ 216 0x07 0x1d 0x16 0x98 LDA r8, [p7], #4
132
+ 220 0x07 0x2e 0x76 0x98 LDA r19, [p7], #8
133
+ 224 0xff 0x93 0x24 0xdd 0x81 0xd4 LDA p1, [sp, #-4]; MOV p2, p7
134
+ 230 0x00 0x00 NOPX
135
+ 232 0x04 0x00 0x27 0x31 0x39 0xe4 MOVX r16, #0; MOV el7, r24
136
+ 238 0x1c 0x90 0x9c 0xf8 MOV el9, r16
137
+ 242 0x1d 0x14 0xa9 0x58 ADD.NC r20, r9, r10
138
+ 246 0x00 0x2c 0xf0 0x98 0x8b 0x02 0x8d 0x10 0xa2 0xba NOPA; MOVS p0, p6; ADD.NC r20, r20, r8
139
+ 256 0x00 0xd2 0x00 0x00 0x20 0x01 0x5b 0x01 0x30 0x0a 0xbc 0xe8 0xa8 0x00 0x00 0xe1 MOVA r18, #6; NOPB; NOPS; MOVX r19, #0; ADD.NC lc, r19, r20; NOPV
140
+ .label ZLS_F_main_272
141
+ .loop_nesting 3
142
+ .begin_of_loop
143
+ 272 0x43 0xe4 0xd1 0x1d 0xe9 0x82 0x6c 0xc0 0x42 0xba LDA dn6, [p2], #4; ST el7, [p1], #4; ADD.NC r19, r19, #1
144
+ 282 0x43 0xd2 0xd9 0x32 0x04 0x14 LDA r20, [p2], #4; ADD.NC r18, r18, #4
145
+ 288 0x02 0x1f 0xe6 0x98 LDA dc7, [p2], #4
146
+ 292 0x02 0x1e 0xee 0x98 LDA el11, [p2], #4
147
+ 296 0x00 0x00 NOPX
148
+ 298 0x00 0x00 NOPX
149
+ 300 0x00 0x00 NOPX
150
+ 302 0x09 0x1f 0x21 0x98 ST dn6, [p1], #4
151
+ 306 0x00 0x00 NOPX
152
+ 308 0x09 0x1f 0xe1 0x98 ST dc7, [p1], #4
153
+ 312 0x23 0xcd 0x30 0x00 0x01 0xa5 0x70 0x02 ST el9, [p1], #4; NOPM
154
+ 320 0x00 0x2c 0xf0 0x00 0x21 0x1e 0x91 0x80 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; ST r20, [p1], #4; NOPX; NOPM; NOPV
155
+ 336 0x00 0x2c 0xf0 0x00 0x21 0xbe 0xe9 0x80 0x00 0x00 0x01 0xa5 0x78 0x00 0x00 0xe1 NOPA; NOPB; ST el11, [p1], #-20; NOPX; NOPM; NOPV
156
+ 352 0x00 0x2c 0xf0 0x00 0x20 0x01 0x5b 0x00 0x00 0x02 0xa9 0x60 0x78 0x00 0x00 0xe1 NOPA; NOPB; NOPS; NOPX; MOV r21, p1; NOPV
157
+ .label ZLE_F_main_368
158
+ .end_of_loop
159
+ 368 0x00 0x2c 0xf0 0x00 0x20 0x1e 0xb1 0x80 0x00 0x00 0xb5 0x46 0x08 0x00 0x00 0xe1 NOPA; NOPB; ST r21, [p0], #4; NOPX; ADD.NC p1, r21, #24; NOPV
160
+ .loop_nesting 2
161
+ 384 0x00 0x0b 0x00 0x27 0x46 0x6e 0x6f 0x60 0x78 0xba MOVA r11, #0; LSHL r20, r19, r12; MOV r19, p7
162
+ 394 0x9b 0x74 0x69 0xb2 0x01 0x24 ADD r13, r19, #-24; ADD.NC r19, r18, #1
163
+ 400 0xf1 0xa1 0x60 0x25 0x26 0x6f 0x45 0x10 0x79 0x3a MOVS p7, r13; LSHL r18, r18, r12; MOV dj6, r20
164
+ 410 0x9c 0x99 0xbb 0x12 0x41 0xe4 LSHL r18, r19, r12; MOV dj5, r18
165
+ 416 0xf4 0x4a 0xde 0x8d 0x92 0x94 LDA r18, [p7, dj5]; ADD.NC dn7, r13, r18
166
+ 422 0xd8 0x74 0x30 0x02 0x2c 0x7f 0xc0 0x02 ST dn7, [p6, dj6]; ADD.NC r17, r17, #-1
167
+ 430 0x00 0x00 NOPX
168
+ 432 0x00 0x00 NOPX
169
+ 434 0x00 0x00 NOPX
170
+ 436 0x00 0x00 NOPX
171
+ 438 0x00 0x00 NOPX
172
+ 440 0x00 0x2b 0x60 0x02 0x4c 0xe4 0xa0 0x02 NOPS; ADD.NC r18, r19, r18
173
+ .label TGT_F_main_448
174
+ .loop_nesting 3
175
+ 448 0x14 0xa1 0x01 0x98 SUB r16, r18, r16
176
+ 452 0x14 0x1e 0xcd 0x98 LSHL r15, r16, r12
177
+ 456 0xf1 0xa1 0x60 0x03 0xc3 0xd0 0x70 0x02 MOVS p7, r13; MOV dj7, r15
178
+ 464 0x07 0xe0 0x16 0x98 LDA r0, [p7, dj7]
179
+ .no_stack_arguments
180
+ 468 0x00 0x00 0x00 0x00 0x01 0x04 JL #_ZN3adf11block_writeEPKNS_7reg_valEj
181
+ .delay_slot
182
+ 474 0x1b 0x98 0x00 0x98 ADD.NC r14, r16, #1
183
+ .delay_slot
184
+ .swstall delay_slot
185
+ 478 0x00 0x00 NOPX
186
+ .delay_slot
187
+ 480 0x13 0xa0 0xcd 0x98 LSHL r16, r14, r12
188
+ .delay_slot
189
+ 484 0x18 0x66 0xc1 0x58 ADD.NC p0, r13, r16
190
+ .delay_slot
191
+ 488 0xfe 0x46 0xb0 0x00 0x01 0xa5 0x70 0x02 ST r17, [sp, #-16]; NOPM
192
+ .return_address
193
+ 496 0x04 0x40 0xa1 0x0f 0x41 0xe4 MOVX r17, #1; MOV dj0, r15
194
+ 502 0xe0 0x4a 0xd0 0x00 0x00 0x1e 0x8b 0xe0 0x10 0xba LDA r18, [p7, dj0]; MOVXM r20, #30656
195
+ 512 0x00 0x00 0x7a 0xaf 0x84 0x44 MOVXM r21, #30658
196
+ 518 0x00 0x08 0x00 0xc0 0x00 0x44 MOVXM p0, #524288
197
+ 524 0x00 0x70 0x08 0x20 0x06 0x44 MOVXM r16, #7340035
198
+ 530 0x00 0x00 NOPX
199
+ 532 0x00 0x00 NOPX
200
+ 534 0x00 0x00 NOPX
201
+ 536 0x14 0xa3 0x1d 0x98 LSHL r17, r18, r17
202
+ 540 0x1c 0x98 0xb9 0x58 ADD.NC r18, r17, r14
203
+ 544 0x14 0xa6 0xcd 0x98 LSHL r19, r18, r12
204
+ 548 0x18 0x89 0xa0 0xf8 MOV dj0, r19
205
+ 552 0x07 0x02 0xd6 0x98 LDA r22, [p7, dj0]
206
+ 556 0x00 0x00 NOPX
207
+ 558 0x00 0x00 NOPX
208
+ 560 0x00 0x00 NOPX
209
+ 562 0x00 0x00 NOPX
210
+ 564 0x00 0x00 NOPX
211
+ 566 0x00 0x00 NOPX
212
+ 568 0x15 0xb6 0xcc 0x98 LTU r27, r22, r12
213
+ 572 0x15 0x69 0x42 0x18 SEL.EQZ r20, r21, r20, r27
214
+ 576 0x1d 0x1b 0x51 0x58 ADD.NC r20, r22, r20
215
+ 580 0x15 0x28 0xcd 0x98 LSHL r20, r20, r12
216
+ 584 0x00 0x2b 0x60 0x00 0x45 0x10 0x70 0x02 NOPS; MOV dj0, r20
217
+ .label TGT_F_main_592
218
+ .loop_nesting 4
219
+ 592 0x00 0x02 0x93 0x98 LDA.TM r20, [p0, dj0]
220
+ 596 0x00 0x00 NOPX
221
+ 598 0x00 0x00 NOPX
222
+ 600 0x00 0x00 NOPX
223
+ 602 0x00 0x00 NOPX
224
+ 604 0x00 0x00 NOPX
225
+ 606 0x00 0x00 NOPX
226
+ 608 0x15 0x2b 0x04 0x98 AND r21, r20, r16
227
+ 612 0xa8 0x01 0x28 0x40 0x01 0x84 JNZ r21, #TGT_F_main_592
228
+ .delay_slot
229
+ .swstall delay_slot
230
+ 618 0x00 0x00 NOPX
231
+ .delay_slot
232
+ .swstall delay_slot
233
+ 620 0x00 0x00 NOPX
234
+ .delay_slot
235
+ .swstall delay_slot
236
+ 622 0x00 0x00 NOPX
237
+ .delay_slot
238
+ .swstall delay_slot
239
+ 624 0x00 0x00 NOPX
240
+ .delay_slot
241
+ .swstall delay_slot
242
+ 626 0x00 0x00 NOPX
243
+ .loop_nesting 3
244
+ 628 0x1b 0xd9 0x00 0x98 ADD.NC r15, r18, #1
245
+ 632 0x13 0xdc 0xcd 0x98 LSHL r14, r15, r12
246
+ 636 0x19 0x87 0x20 0xf8 MOV dj1, r14
247
+ 640 0x07 0x20 0x16 0x98 LDA r0, [p7, dj1]
248
+ .no_stack_arguments
249
+ 644 0x00 0x00 0x00 0x00 0x01 0x04 JL #_ZN3adf11block_writeEPKNS_7reg_valEj
250
+ .delay_slot
251
+ .swstall delay_slot
252
+ 650 0x00 0x00 NOPX
253
+ .delay_slot
254
+ .swstall delay_slot
255
+ 652 0x00 0x00 NOPX
256
+ .delay_slot
257
+ 654 0x13 0x5a 0x23 0x18 ADD r13, r13, #8
258
+ .delay_slot
259
+ 658 0x18 0x69 0xb5 0x58 ADD.NC p0, r19, r13
260
+ .delay_slot
261
+ 662 0x00 0x2c 0xf7 0xea 0x35 0x80 0x00 0x00 0x00 0x7a NOPA; ST r17, [sp, #-24]; NOPX
262
+ .return_address
263
+ 672 0x00 0x0e 0x00 0x01 0x00 0x28 0x43 0x90 0x78 0xba MOVA r14, #0; MOVX r16, #1; MOV dj0, r14
264
+ 682 0xe0 0x4a 0xd8 0xad 0xfc 0x14 LDA r18, [p7, dj0]; ADD.NC r17, r13, #-4
265
+ 688 0x0f 0xee 0x35 0x98 ST r17, [sp, #-20]
266
+ 692 0x00 0x00 NOPX
267
+ 694 0x00 0x00 NOPX
268
+ 696 0x00 0x00 NOPX
269
+ 698 0x00 0x00 NOPX
270
+ 700 0x00 0x00 NOPX
271
+ 702 0x14 0xa5 0x0d 0x98 LSHL r18, r18, r16
272
+ 706 0x14 0xa1 0x05 0x98 OR r16, r18, r16
273
+ 710 0x1c 0x97 0xc1 0x58 ADD.NC r18, r15, r16
274
+ 714 0xfc 0x42 0xb0 0x24 0xc6 0x6c 0x37 0x60 0x79 0x3a ST r16, [sp, #-32]; LSHL r12, r18, r12; MOV p0, p7
275
+ 724 0xfc 0xca 0xb0 0x23 0x06 0x04 0x43 0x10 0x79 0x3a ST r18, [sp, #-28]; ADD r16, r17, r12; MOV dj0, r12
276
+ 734 0x00 0x42 0xd7 0xde 0x15 0x80 0x00 0x03 0xb1 0x80 0x10 0x76 LDA r16, [p0, dj0]; ST r16, [sp, #-36]; MOVXM p7, #(TGT_F_main_768 + 0)
277
+ 746 0x00 0x00 NOPX
278
+ 748 0x00 0x00 NOPX
279
+ 750 0x00 0x00 NOPX
280
+ 752 0x00 0x00 NOPX
281
+ 754 0x00 0x00 NOPX
282
+ 756 0x00 0x00 NOPX
283
+ 758 0x00 0x2c 0xf0 0x00 0x10 0x01 0xec 0x3f 0xce 0xba NOPA; NOPB; ADD.NC r15, r16, #-1
284
+ .label TGT_F_main_768
285
+ .loop_nesting 4
286
+ 768 0x12 0xe0 0xe5 0x98 OR r16, r11, r14
287
+ 772 0x80 0x01 0xb8 0x40 0x01 0x84 JNZ r16, #TGT_F_main_880
288
+ .delay_slot
289
+ .swstall delay_slot
290
+ 778 0x00 0x00 NOPX
291
+ .delay_slot
292
+ .swstall delay_slot
293
+ 780 0x00 0x00 NOPX
294
+ .delay_slot
295
+ .swstall delay_slot
296
+ 782 0x00 0x00 NOPX
297
+ .delay_slot
298
+ .swstall delay_slot
299
+ 784 0x00 0x00 NOPX
300
+ .delay_slot
301
+ 786 0x11 0xa1 0x60 0x00 0xc3 0x10 0x70 0x02 MOVS p0, r13; MOV dj1, r12
302
+ 794 0x07 0xdc 0x99 0x18 LDA p1, [sp, #-36]
303
+ 798 0x00 0x00 NOPX
304
+ 800 0x00 0x00 NOPX
305
+ 802 0x00 0x00 NOPX
306
+ 804 0x00 0x00 NOPX
307
+ 806 0x00 0x00 NOPX
308
+ 808 0x00 0x00 NOPX
309
+ 810 0x01 0x06 0x16 0x98 LDA r16, [p1]
310
+ 814 0x00 0x00 NOPX
311
+ 816 0x00 0x00 NOPX
312
+ 818 0x00 0x00 NOPX
313
+ 820 0x00 0x00 NOPX
314
+ 822 0x00 0x00 NOPX
315
+ 824 0x00 0x00 NOPX
316
+ 826 0x80 0x01 0xb8 0x00 0x01 0x84 JZ r16, #TGT_F_main_880
317
+ .delay_slot
318
+ .swstall delay_slot
319
+ 832 0x00 0x00 NOPX
320
+ .delay_slot
321
+ .swstall delay_slot
322
+ 834 0x00 0x00 NOPX
323
+ .delay_slot
324
+ .swstall delay_slot
325
+ 836 0x00 0x00 NOPX
326
+ .delay_slot
327
+ .swstall delay_slot
328
+ 838 0x00 0x00 NOPX
329
+ .delay_slot
330
+ .swstall delay_slot
331
+ 840 0x00 0x00 NOPX
332
+ .swstall chess_separator_scheduler
333
+ 842 0x00 0x00 NOPX
334
+ .swstall chess_separator_scheduler
335
+ 844 0x00 0x00 NOPX
336
+ .swstall chess_separator_scheduler
337
+ 846 0x00 0x00 NOPX
338
+ .swstall chess_separator_scheduler
339
+ 848 0x00 0x00 NOPX
340
+ .swstall chess_separator_scheduler
341
+ 850 0x00 0x00 NOPX
342
+ .swstall chess_separator_scheduler
343
+ 852 0x00 0x00 NOPX
344
+ 854 0x10 0x08 0x00 0x18 DONE
345
+ .swstall chess_separator_scheduler
346
+ 858 0x00 0x00 NOPX
347
+ .swstall chess_separator_scheduler
348
+ 860 0x00 0x00 NOPX
349
+ .swstall chess_separator_scheduler
350
+ 862 0x00 0x00 NOPX
351
+ .swstall chess_separator_scheduler
352
+ 864 0x00 0x00 NOPX
353
+ .swstall chess_separator_scheduler
354
+ 866 0x00 0x00 NOPX
355
+ .swstall chess_separator_scheduler
356
+ 868 0x00 0x2c 0xf0 0x00 0x20 0x00 0x00 0x00 0x00 0xad 0x81 0x36 NOPA; NOPB; NOPS; NOPX
357
+ .label TGT_F_main_880
358
+ 880 0x04 0x02 0xd0 0xd9 0x81 0xd4 LDA r0, [p0, dj1]; MOV p0, p6
359
+ .no_stack_arguments
360
+ 886 0x00 0x00 0x00 0x00 0x01 0x04 JL #_Z13kernelWrapperPPvjjjj
361
+ .delay_slot
362
+ .swstall delay_slot
363
+ 892 0x00 0x00 NOPX
364
+ .delay_slot
365
+ .swstall delay_slot
366
+ 894 0x00 0x00 NOPX
367
+ .delay_slot
368
+ 896 0x18 0x55 0x20 0xf8 MOV r1, r10
369
+ .delay_slot
370
+ 900 0x18 0x94 0xa0 0xf8 MOV r2, r9
371
+ .delay_slot
372
+ 904 0x00 0x2b 0x60 0x00 0x6a 0x10 0x70 0x02 NOPS; MOV r3, r8
373
+ .return_address
374
+ 912 0x13 0xdf 0xe0 0x18 JNZD r15, r15, p7
375
+ .delay_slot
376
+ .swstall delay_slot
377
+ 916 0x00 0x00 NOPX
378
+ .delay_slot
379
+ .swstall delay_slot
380
+ 918 0x00 0x00 NOPX
381
+ .delay_slot
382
+ .swstall delay_slot
383
+ 920 0x00 0x00 NOPX
384
+ .delay_slot
385
+ .swstall delay_slot
386
+ 922 0x00 0x00 NOPX
387
+ .delay_slot
388
+ 924 0x1b 0x97 0x00 0x98 ADD.NC r14, r14, #1
389
+ .loop_nesting 3
390
+ 928 0xfe 0x46 0x20 0x00 0xc0 0x49 0x6a 0xc0 0x48 0xba LDA r17, [sp, #-16]; MOVX r12, #2; ADD.NC r11, r11, #1
391
+ 938 0xfd 0xca 0x20 0x00 0x00 0x03 0xb0 0xe0 0x10 0xba LDA r18, [sp, #-20]; MOVXM p7, #(TGT_F_main_448 + 0)
392
+ 948 0x07 0xe6 0x71 0x18 LDA r19, [sp, #-28]
393
+ 952 0x07 0xe2 0x11 0x18 LDA r16, [sp, #-32]
394
+ 956 0x07 0xea 0x91 0x18 LDA r20, [sp, #-24]
395
+ 960 0x00 0x00 NOPX
396
+ 962 0x00 0x00 NOPX
397
+ 964 0x14 0x63 0xe0 0x18 JNZD r17, r17, p7
398
+ .delay_slot
399
+ 968 0x14 0x9b 0xf3 0x18 ADD r13, r18, #-4
400
+ .delay_slot
401
+ 972 0x1c 0x99 0x81 0x98 ADD.NC r18, r19, #3
402
+ .delay_slot
403
+ .swstall delay_slot
404
+ 976 0x00 0x00 NOPX
405
+ .delay_slot
406
+ 978 0x1c 0x1a 0x41 0x58 ADD.NC r16, r20, r16
407
+ .delay_slot
408
+ 982 0x1c 0x18 0x02 0x98 ADD.NC r16, r16, #5
409
+ .loop_nesting 2
410
+ 986 0x07 0xd2 0x07 0x8d 0x0b 0x25 0x06 0x6e 0x88 0x3f 0x58 0x76 MOVA r18, #62; MOVS p7, r13; LSHL r16, r18, r12; MOV r20, #63
411
+ 998 0x00 0x30 0x00 0x3f 0x67 0xe8 0x44 0x10 0x78 0xba MOVA r16, #1; MOVX r22, #-1; MOV dj0, r16
412
+ 1008 0xe0 0x46 0xd0 0x00 0x00 0x02 0x68 0x00 0x10 0xba LDA r17, [p7, dj0]; MOVXM r19, #lcpPing
413
+ 1018 0x9e 0xda 0xfc 0x20 0x01 0x64 EQ r27, r19, r13; MOV r24, #0
414
+ 1024 0x15 0x2f 0x22 0x18 SEL.EQZ r23, r20, r18, r27
415
+ 1028 0x00 0x00 0x0a 0xa0 0x00 0x44 MOVXM r21, #lcpPong
416
+ 1034 0x00 0x00 NOPX
417
+ 1036 0x00 0x00 NOPX
418
+ 1038 0x00 0x00 NOPX
419
+ 1040 0x14 0x74 0xd0 0x18 EQZ r26, r17
420
+ 1044 0x15 0xd1 0x68 0x18 REL r23, r22
421
+ 1048 0x14 0xa5 0x42 0x18 SEL.EQZ r18, r18, r20, r27
422
+ 1052 0x00 0x00 NOPX
423
+ 1054 0x00 0x00 NOPX
424
+ 1056 0x14 0x97 0x08 0x18 ACQ.COND r18, r16, r26
425
+ 1060 0x88 0x00 0x60 0x00 0x01 0x84 JZ r17, #TGT_F_main_192
426
+ .delay_slot
427
+ 1066 0x14 0xe1 0x52 0x18 SEL.EQZ r16, r19, r21, r27
428
+ .delay_slot
429
+ 1070 0x1f 0x68 0x20 0xf8 MOV p7, r16
430
+ .delay_slot
431
+ .swstall delay_slot
432
+ 1074 0x00 0x00 NOPX
433
+ .delay_slot
434
+ .swstall delay_slot
435
+ 1076 0x00 0x00 NOPX
436
+ .delay_slot
437
+ .swstall delay_slot
438
+ 1078 0x00 0x00 NOPX
439
+ .loop_nesting 1
440
+ 1080 0xff 0x03 0x20 0x01 0x90 0x0a 0x08 0x01 0x58 0xba LDA p0, [sp, #-8]; MOVX r25, #0; MOV r16, #1
441
+ 1090 0x00 0x00 0x18 0x00 0x00 0x84 J #TGT_F_main_48
442
+ .delay_slot
443
+ .swstall delay_slot
444
+ 1096 0x00 0x00 NOPX
445
+ .delay_slot
446
+ .swstall delay_slot
447
+ 1098 0x00 0x00 NOPX
448
+ .delay_slot
449
+ .swstall delay_slot
450
+ 1100 0x00 0x00 NOPX
451
+ .delay_slot
452
+ .swstall delay_slot
453
+ 1102 0x00 0x00 NOPX
454
+ .label _main__end last
455
+ .label _main___func_end0 last
456
+ .delay_slot
457
+ 1104 0x07 0xf4 0x99 0x18 LDA p1, [sp, #-12]
458
+
459
+ .undef global data main
460
+
461
+ .undef global data lcpPing
462
+
463
+ .undef global data lcpPong
464
+
465
+ .undef global text _ZN3adf11block_writeEPKNS_7reg_valEj
466
+
467
+ .undef global text _Z13kernelWrapperPPvjjjj
468
+
469
+
470
+
471
+ .direct_eval +,-,(,)
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/Release/0_0.sdr ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ // File generated by bridge version V-2024.06#84922c0d9f#241219, Fri Mar 21 03:42:51 2025
3
+ // Copyright 2014-2024 Synopsys, Inc. All rights reserved.
4
+ // bridge -o../Release/0_0 ../Release/0_0.o -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/isg -g -I/usr/local/lib/python3.10/dist-packages/include -I/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend -I/usr/local/lib/python3.10/site-packages/include/aie_api -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common -I/usr/local/lib/python3.10/dist-packages/vitis_mllib -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc -I/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf -I/app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/. -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/include -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libs/libcxx-9.0.0/include-lite -I/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/include -D__AIENGINE__ -D__AIE_ARCH__=21 -DDEPLOYMENT_ELF=1 -D__LOCK_FENCE_MODE__=0 -D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ -DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1 -DAIE_OPTION_SCALAR_FLOAT_ON_VECTOR -D__tct_tgt__=241219 -c0_0.bcf -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/softfloat/lib/Release -L/usr/local/lib/python3.10/dist-packages/data/aie2p/lib/runtime_cxx/libcxx-lite/lib/Release_LLVM -lme -lc -lm -lc++lite -lsoftfloat -S -export-locals -iconfig extra_memories.bcf -yTM -m -fC -fS -fH +m -T +work ../Release/chesswork848 -pme
5
+
6
+ // Release: ipp V-2024.06-TGT-241219
7
+
8
+ // Symbols in memory 'DM_bankA':
9
+ // Symbols in memory 'DM_bankAB':
10
+ // Symbols in memory 'DM_bankAC':
11
+ // Symbols in memory 'DM_bankAD':
12
+ // Symbols in memory 'DM_bankB':
13
+ // Symbols in memory 'DM_bankBC':
14
+ // Symbols in memory 'DM_bankBD':
15
+ // Symbols in memory 'DM_bankC':
16
+ // Symbols in memory 'DM_bankCD':
17
+ // Symbols in memory 'DM_bankD':
18
+ // Symbols in memory 'DM_stack':
19
+ // Symbols in memory 'DM_test':
20
+ // Symbols in memory 'DMb':
21
+ _symbol lcpPing 0x0007b280
22
+ _symbol lcpPong 0x0007b680
23
+ // Symbols in memory 'DMh':
24
+ // Symbols in memory 'DMh_bankA':
25
+ // Symbols in memory 'DMh_bankAB':
26
+ // Symbols in memory 'DMh_bankAC':
27
+ // Symbols in memory 'DMh_bankAD':
28
+ // Symbols in memory 'DMh_bankB':
29
+ // Symbols in memory 'DMh_bankBC':
30
+ // Symbols in memory 'DMh_bankBD':
31
+ // Symbols in memory 'DMh_bankC':
32
+ // Symbols in memory 'DMh_bankCD':
33
+ // Symbols in memory 'DMh_bankD':
34
+ // Symbols in memory 'DMh_stack':
35
+ // Symbols in memory 'DMs':
36
+ // Symbols in memory 'DMs_bankA':
37
+ // Symbols in memory 'DMs_bankAB':
38
+ // Symbols in memory 'DMs_bankAC':
39
+ // Symbols in memory 'DMs_bankAD':
40
+ // Symbols in memory 'DMs_bankB':
41
+ // Symbols in memory 'DMs_bankBC':
42
+ // Symbols in memory 'DMs_bankBD':
43
+ // Symbols in memory 'DMs_bankC':
44
+ // Symbols in memory 'DMs_bankCD':
45
+ // Symbols in memory 'DMs_bankD':
46
+ // Symbols in memory 'DMs_stack':
47
+ // Symbols in memory 'DMv':
48
+ // Symbols in memory 'DMv_bankA':
49
+ // Symbols in memory 'DMv_bankAB':
50
+ // Symbols in memory 'DMv_bankAC':
51
+ // Symbols in memory 'DMv_bankAD':
52
+ // Symbols in memory 'DMv_bankB':
53
+ // Symbols in memory 'DMv_bankBC':
54
+ // Symbols in memory 'DMv_bankBD':
55
+ // Symbols in memory 'DMv_bankC':
56
+ // Symbols in memory 'DMv_bankCD':
57
+ // Symbols in memory 'DMv_bankD':
58
+ // Symbols in memory 'DMv_stack':
59
+ // Symbols in memory 'DMw':
60
+ // Symbols in memory 'DMw_bankA':
61
+ // Symbols in memory 'DMw_bankAB':
62
+ // Symbols in memory 'DMw_bankAC':
63
+ // Symbols in memory 'DMw_bankAD':
64
+ // Symbols in memory 'DMw_bankB':
65
+ // Symbols in memory 'DMw_bankBC':
66
+ // Symbols in memory 'DMw_bankBD':
67
+ // Symbols in memory 'DMw_bankC':
68
+ // Symbols in memory 'DMw_bankCD':
69
+ // Symbols in memory 'DMw_bankD':
70
+ // Symbols in memory 'DMw_stack':
71
+ // Symbols in memory 'DMx':
72
+ // Symbols in memory 'DMx_bankA':
73
+ // Symbols in memory 'DMx_bankAB':
74
+ // Symbols in memory 'DMx_bankAC':
75
+ // Symbols in memory 'DMx_bankAD':
76
+ // Symbols in memory 'DMx_bankB':
77
+ // Symbols in memory 'DMx_bankBC':
78
+ // Symbols in memory 'DMx_bankBD':
79
+ // Symbols in memory 'DMx_bankC':
80
+ // Symbols in memory 'DMx_bankCD':
81
+ // Symbols in memory 'DMx_bankD':
82
+ // Symbols in memory 'DMx_stack':
83
+ // Symbols in memory 'PM':
84
+ _symbol _main_init 0x00000000
85
+ _symbol _main 0x000000e0
86
+ _symbol _ZN3adf11block_writeEPKNS_7reg_valEj 0x00000540
87
+ _symbol _fini 0x000005f0
88
+ _symbol __cxa_finalize 0x000006a0
89
+ // Symbols in memory 'PMw':
90
+ // Symbols in memory 'TM4':
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/elf_ctrl_pkt.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7bae6d17370fa6ff9dce1f15ddab525892dba32c85b49d7303d98c7938009da9
3
+ size 6148
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_0.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:981b0e1230eb5dbd26bf35b2717875f8d4f352bc7974fd48c2ee0d81ca343686
3
+ size 296
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_1.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e06471c10dc9f1b825c878a76d2d029bada052ace379236c9916a0d302d6e270
3
+ size 300
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_100.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ae43a8dcc0cd99130f42249d4f2a9f1fc91c3c99699db65bb98ad360a326ef6
3
+ size 376
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/lcp/0_0_11.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:05761cae01b301200d2d2ef4aa2f3f81493591bd9ea0b012b956cd6fef12f5dd
3
+ size 384
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/0_0/xlopt.log ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ERROR: ld.so: object '/lib64/libudev.so.1' from LD_PRELOAD cannot be preloaded (cannot open shared object file): ignored.
2
+
3
+
4
+ --------------- FILTER ANALYSIS INFO LOG ---------------
5
+
6
+ Reading Header IR from ir/_header.ll
7
+
8
+ --------------------------------------------------------
9
+
10
+ Add module pass 363{anonymous}::GuidancePass
11
+ Add module pass 363{anonymous}::ChessOptionsPass
12
+ Add module pass 363{anonymous}::DisableInliningInMainPass
13
+ Add module pass 363cdno::xlopt::AIEMergeSubWordStoresOpt
14
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::PromotePass>
15
+ Add module pass 363{anonymous}::IpConstPropPass
16
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::SimplifyCFGPass>
17
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::LoopSimplifyPass>
18
+ Add module pass 363{anonymous}::XLModuleAdaptor<llvm::FunctionToLoopPassAdaptor>
19
+ Add module pass 363{anonymous}::XLModuleAdaptor<{anonymous}::AIELoopInfoPass>
20
+ Add module pass 363cdno::xlopt::AIEAnnotatePragmaPass
21
+ Add module pass 363{anonymous}::XLModuleAdaptor<{anonymous}::AIELoopPeelPass>
22
+ Add module pass 363{anonymous}::AIEAliasAnalysisPass
23
+
24
+
25
+ --------------- MEMORY MANAGEMENT GUIDANCE LOG ---------------
26
+
27
+ SIZE_HINT: Global array lcpPing is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
28
+ SIZE_HINT: Global array lcpPong is 1024 bytes. Consider making it mapper-managed LUT or memory buffer.
29
+ GLOBAL_RBW_HINT: Global variable 'lcpPong' reads its default/external initialization at 0_0/src/0_0.cc:81 in kernel 'main', and is not explicitly written before this read
30
+ GLOBAL_CONFLICT_HINT: Kernels 'block_write' and 'initialize_lock' have no execution dependence, but access the same global variable 'TM_Address_Space_Start', and one of those accesses is a write
31
+ GLOBAL_CONFLICT_HINT: Kernels 'block_write' and 'wait_dma_channel_done' have no execution dependence, but access the same global variable 'TM_Address_Space_Start', and one of those accesses is a write
32
+ GLOBAL_CONFLICT_HINT: Kernels 'initialize_lock' and 'wait_dma_channel_done' have no execution dependence, but access the same global variable 'TM_Address_Space_Start', and one of those accesses is a write
33
+
34
+ --------------------------------------------------------------
35
+
36
+
37
+
38
+ --------------- MERGING SUBWORD STORES OPT LOG ---------------
39
+
40
+
41
+
42
+ --------------- LOOP STATISTICS : main ---------------
43
+
44
+ Total loops = 7
45
+ Loops with prepare for pipelining pragma = 0
46
+ Loops with unroll pragma = 0
47
+ Loops with flatten pragma = 0
48
+ Loops with min range pragma = 4
49
+ Loops with max range pragma = 2
50
+ Loops with known trip count = 0
51
+
52
+ ------------------------------------------------------
53
+
54
+
55
+
56
+ --------------- LOOP STATISTICS : _ZN3adf11block_writeEPKNS_7reg_valEj ---------------
57
+
58
+ Total loops = 1
59
+ Loops with prepare for pipelining pragma = 0
60
+ Loops with unroll pragma = 0
61
+ Loops with flatten pragma = 0
62
+ Loops with min range pragma = 0
63
+ Loops with max range pragma = 0
64
+ Loops with known trip count = 0
65
+
66
+ --------------------------------------------------------------------------------------
67
+
68
+
69
+
70
+ --------------- LOOP STATISTICS : _ZN3adf21wait_dma_channel_doneEj ---------------
71
+
72
+ Total loops = 1
73
+ Loops with prepare for pipelining pragma = 0
74
+ Loops with unroll pragma = 0
75
+ Loops with flatten pragma = 0
76
+ Loops with min range pragma = 0
77
+ Loops with max range pragma = 0
78
+ Loops with known trip count = 0
79
+
80
+ ----------------------------------------------------------------------------------
81
+
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/AddressSpace.txt ADDED
File without changes
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/AliasAnalysisReport.txt ADDED
File without changes
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/Makefile ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/aie/layer_control_parameters.json ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/Work/top.aiecompile_summary ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/aie.mk ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ aiecompile:
2
+ aiecompiler /app/vaiml_1.3_examples/camo/segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend/top.cpp --part=xc10AIE2P_ML-die-0x-e-S-es1 --nodot-graph --runtime-opt=1 --disable-multirate-analysis --enable-core-processor-bus --enable-multi-layer --heapsize=1792 --stacksize=1400 --max-layer-ctrl-param-size=256 --compile-for-aiesim=false --workdir=Work --multi-layer-ctrl-pkt --aie2ipu-base-addr=0 -enable-light-cdo --Xelfgen=-j4 --multi-layer-pipelining --multi-layer-opt=3 --Xpreproc=-D__IO_BUFFER_FORCE_LIGHT_WEIGHT__ --multi-layer-ext-buf-file=/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend/flexmlrt-hsi.json --enable-partition=0:4 --multi-layer-ctrl-pkt-column-span=4 --multi-layer-prebuilt-archive=/usr/local/lib/python3.10/dist-packages/flexml/flexml_extras/data/ryzen-ai/stx/unified-overlay-4x4.json --multi-layer-prebuilt-archive-enable-elf-gen --multi-layer-init-core-elf-ctrl-pkt --multi-layer-pm-id 29006 --Xpreproc="-DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16=1" --include=/app/vaiml_1.3_examples/camo/./segmentation_1_4_0_fp32_combined/vaiml_par_0/0/backend --include=/usr/local/lib/python3.10/site-packages/include/aie_api --include=/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/common --include=/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/include/common --include=/usr/local/lib/python3.10/dist-packages/vitis_mllib --include=/usr/local/lib/python3.10/dist-packages/vitis_mllib/L1/include/misc --include=/usr/local/lib/python3.10/dist-packages/vitis_mllib/L2/src/ml_adf --output-archive libadf.a --adf-api-log-level=0 --multi-layer-pm-reloading=1 --Xelfgen=-j1
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/aiecompiler-flexml.log ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/aiesim_cmd_opts.txt ADDED
File without changes
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/aiecompiler/xcd.log ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Fri Mar 21 03:30:56 2025: Running xcd server. This server was built by gradle.
2
+ Fri Mar 21 03:30:56 2025: Server was asked to start on port: '34917'
3
+ Fri Mar 21 03:30:56 2025: Server is using token UUID: 'e14f6479-c44a-42cf-8f07-308e6ee9c741'
4
+ Fri Mar 21 03:30:56 2025: SESSION_MGR: creating session with session uuid: '6d04a3e9-ade5-4ab9-8d63-1f78fcd2ec76' and name 'default'
5
+ Fri Mar 21 03:30:56 2025: Attempting to start server on port '34917'
6
+ Fri Mar 21 03:30:56 2025: XCD main fifo: created main fifo '/tmp/xcdmaine14f6479-c44a-42cf-8f07-308e6ee9c741', fd read = 12
7
+ Fri Mar 21 03:30:56 2025: Running Dispatch Server
8
+ Fri Mar 21 03:30:56 2025: Version 2.2.0
9
+ Fri Mar 21 03:30:57 2025: Dispatch Server: Accepted socket connection from client
10
+ Fri Mar 21 03:30:57 2025: Starting Socket connection
11
+ Fri Mar 21 03:30:57 2025: EXCHANGE_TOKEN received, server token: e14f6479-c44a-42cf-8f07-308e6ee9c741, passed token: e14f6479-c44a-42cf-8f07-308e6ee9c741
12
+ Fri Mar 21 03:30:57 2025: SESSION_MGR: creating session with session uuid: 'bef5681c-a6ee-487a-8b13-f1c9f1805b32' and name 'AIE Compilation'
13
+ Fri Mar 21 03:30:57 2025: KERNEL_SVC: creating session with uuid: 'bef5681c-a6ee-487a-8b13-f1c9f1805b32' and name 'AIE Compilation'
14
+ Fri Mar 21 03:30:57 2025: GUIDANCE_SMGR: creating guidance session with uuid: 'bef5681c-a6ee-487a-8b13-f1c9f1805b32' and name 'AIE Compilation'
15
+ Fri Mar 21 03:31:54 2025: Dispatch Server: Accepted socket connection from client
16
+ Fri Mar 21 03:31:54 2025: Starting Socket connection
17
+ Fri Mar 21 04:04:52 2025: SESSION_MGR: closing session for UUID 'bef5681c-a6ee-487a-8b13-f1c9f1805b32' with name ''
18
+ Fri Mar 21 04:04:52 2025: KERNEL_SVC: closing session for UUID 'bef5681c-a6ee-487a-8b13-f1c9f1805b32' with name ''
19
+ Fri Mar 21 04:04:52 2025: GUIDANCE_SMGR: closing session for UUID 'bef5681c-a6ee-487a-8b13-f1c9f1805b32' with name ''
20
+ Fri Mar 21 04:04:52 2025: STOP_SERVER received, server token: e14f6479-c44a-42cf-8f07-308e6ee9c741, passed token: e14f6479-c44a-42cf-8f07-308e6ee9c741
21
+ Fri Mar 21 04:04:52 2025: Socket received request to stop server.
22
+ Fri Mar 21 04:04:52 2025: Dispatch Server: do_await_stop
23
+ Fri Mar 21 04:04:52 2025: Dispatch Server: stopping mainFifo_
24
+ Fri Mar 21 04:04:52 2025: XCD main fifo: Stopping
25
+ Fri Mar 21 04:04:52 2025: XCD main fifo: do_read error code: Operation canceled
26
+ Fri Mar 21 04:04:52 2025: XCD main fifo: bytes_transferred: 0
27
+ Fri Mar 21 04:04:52 2025: SERVER: destructor for server bound to port 34917
28
+ Fri Mar 21 04:04:52 2025: XCD main fifo: Good unlink of /tmp/xcdmaine14f6479-c44a-42cf-8f07-308e6ee9c741
29
+ Fri Mar 21 04:04:52 2025: Server exiting with status 0.
30
+ Fri Mar 21 04:04:52 2025: Closing log file.
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/buffer_info.json ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/ctrlPktPatch.json ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/external_buffer_id.json ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/flexmlrt-hsi.json ADDED
@@ -0,0 +1,1253 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "inputs" : [
3
+ {
4
+ "name" : "compute_graph.ifm_ddr",
5
+ "scale_factor" : -1,
6
+ "cpu_shape" : [
7
+ 1,
8
+ 180,
9
+ 320,
10
+ 4
11
+ ],
12
+ "cpu_format" : "NCHW",
13
+ "cpu_dtype" : "fp32",
14
+ "hw_shape" : [
15
+ 320,
16
+ 23,
17
+ 4,
18
+ 1,
19
+ 8
20
+ ],
21
+ "hw_format" : "HCWNC8",
22
+ "hw_dtype" : "bf16"
23
+ },
24
+ {
25
+ "name" : "compute_graph.ifm_ddr_1",
26
+ "scale_factor" : -1,
27
+ "cpu_shape" : [
28
+ 1,
29
+ 16,
30
+ 90,
31
+ 160
32
+ ],
33
+ "cpu_format" : "NCHW",
34
+ "cpu_dtype" : "fp32",
35
+ "hw_shape" : [
36
+ 90,
37
+ 2,
38
+ 160,
39
+ 1,
40
+ 8
41
+ ],
42
+ "hw_format" : "HCWNC8",
43
+ "hw_dtype" : "bf16"
44
+ },
45
+ {
46
+ "name" : "compute_graph.ifm_ddr_2",
47
+ "scale_factor" : -1,
48
+ "cpu_shape" : [
49
+ 1,
50
+ 20,
51
+ 45,
52
+ 80
53
+ ],
54
+ "cpu_format" : "NCHW",
55
+ "cpu_dtype" : "fp32",
56
+ "hw_shape" : [
57
+ 45,
58
+ 3,
59
+ 80,
60
+ 1,
61
+ 8
62
+ ],
63
+ "hw_format" : "HCWNC8",
64
+ "hw_dtype" : "bf16"
65
+ },
66
+ {
67
+ "name" : "compute_graph.ifm_ddr_3",
68
+ "scale_factor" : -1,
69
+ "cpu_shape" : [
70
+ 1,
71
+ 40,
72
+ 23,
73
+ 40
74
+ ],
75
+ "cpu_format" : "NCHW",
76
+ "cpu_dtype" : "fp32",
77
+ "hw_shape" : [
78
+ 23,
79
+ 5,
80
+ 40,
81
+ 1,
82
+ 8
83
+ ],
84
+ "hw_format" : "HCWNC8",
85
+ "hw_dtype" : "bf16"
86
+ },
87
+ {
88
+ "name" : "compute_graph.ifm_ddr_4",
89
+ "scale_factor" : -1,
90
+ "cpu_shape" : [
91
+ 1,
92
+ 64,
93
+ 12,
94
+ 20
95
+ ],
96
+ "cpu_format" : "NCHW",
97
+ "cpu_dtype" : "fp32",
98
+ "hw_shape" : [
99
+ 12,
100
+ 8,
101
+ 20,
102
+ 1,
103
+ 8
104
+ ],
105
+ "hw_format" : "HCWNC8",
106
+ "hw_dtype" : "bf16"
107
+ }
108
+ ],
109
+ "outputs" : [
110
+ {
111
+ "name" : "compute_graph.ofm_ddr_3_l2l3_291_spill",
112
+ "scale_factor" : -1,
113
+ "cpu_shape" : [
114
+ 1,
115
+ 16,
116
+ 90,
117
+ 160
118
+ ],
119
+ "cpu_format" : "NCHW",
120
+ "cpu_dtype" : "fp32",
121
+ "hw_shape" : [
122
+ 90,
123
+ 2,
124
+ 160,
125
+ 1,
126
+ 8
127
+ ],
128
+ "hw_format" : "HCWNC8",
129
+ "hw_dtype" : "bf16"
130
+ },
131
+ {
132
+ "name" : "compute_graph.ofm_ddr_2_l2l3_272_spill",
133
+ "scale_factor" : -1,
134
+ "cpu_shape" : [
135
+ 1,
136
+ 20,
137
+ 45,
138
+ 80
139
+ ],
140
+ "cpu_format" : "NCHW",
141
+ "cpu_dtype" : "fp32",
142
+ "hw_shape" : [
143
+ 45,
144
+ 3,
145
+ 80,
146
+ 1,
147
+ 8
148
+ ],
149
+ "hw_format" : "HCWNC8",
150
+ "hw_dtype" : "bf16"
151
+ },
152
+ {
153
+ "name" : "compute_graph.ofm_ddr_1_l2l3_252_spill",
154
+ "scale_factor" : -1,
155
+ "cpu_shape" : [
156
+ 1,
157
+ 40,
158
+ 23,
159
+ 40
160
+ ],
161
+ "cpu_format" : "NCHW",
162
+ "cpu_dtype" : "fp32",
163
+ "hw_shape" : [
164
+ 23,
165
+ 5,
166
+ 40,
167
+ 1,
168
+ 8
169
+ ],
170
+ "hw_format" : "HCWNC8",
171
+ "hw_dtype" : "bf16"
172
+ },
173
+ {
174
+ "name" : "compute_graph.ofm_ddr_0_l2l3_229_spill",
175
+ "scale_factor" : -1,
176
+ "cpu_shape" : [
177
+ 1,
178
+ 64,
179
+ 12,
180
+ 20
181
+ ],
182
+ "cpu_format" : "NCHW",
183
+ "cpu_dtype" : "fp32",
184
+ "hw_shape" : [
185
+ 12,
186
+ 8,
187
+ 20,
188
+ 1,
189
+ 8
190
+ ],
191
+ "hw_format" : "HCWNC8",
192
+ "hw_dtype" : "bf16"
193
+ },
194
+ {
195
+ "name" : "compute_graph.ofm_ddr_5",
196
+ "scale_factor" : -1,
197
+ "cpu_shape" : [
198
+ 1,
199
+ 3,
200
+ 180,
201
+ 320
202
+ ],
203
+ "cpu_format" : "NCHW",
204
+ "cpu_dtype" : "fp32",
205
+ "hw_shape" : [
206
+ 180,
207
+ 1,
208
+ 320,
209
+ 1,
210
+ 8
211
+ ],
212
+ "hw_format" : "HCWNC8",
213
+ "hw_dtype" : "bf16"
214
+ },
215
+ {
216
+ "name" : "compute_graph.ofm_ddr_4",
217
+ "scale_factor" : -1,
218
+ "cpu_shape" : [
219
+ 1,
220
+ 1,
221
+ 180,
222
+ 320
223
+ ],
224
+ "cpu_format" : "NCHW",
225
+ "cpu_dtype" : "fp32",
226
+ "hw_shape" : [
227
+ 180,
228
+ 1,
229
+ 320,
230
+ 1,
231
+ 8
232
+ ],
233
+ "hw_format" : "HCWNC8",
234
+ "hw_dtype" : "bf16"
235
+ }
236
+ ],
237
+ "weights" : {
238
+ "layers" : [
239
+ {
240
+ "name" : "compute_graph.Layer_8_wts_ddr",
241
+ "offset" : 0
242
+ },
243
+ {
244
+ "name" : "compute_graph.Layer_13_wts_ddr",
245
+ "offset" : 1216
246
+ },
247
+ {
248
+ "name" : "compute_graph.Layer_14_wts_ddr",
249
+ "offset" : 1472
250
+ },
251
+ {
252
+ "name" : "compute_graph.Layer_15_wts_ddr",
253
+ "offset" : 2560
254
+ },
255
+ {
256
+ "name" : "compute_graph.Layer_16_wts_ddr",
257
+ "offset" : 4736
258
+ },
259
+ {
260
+ "name" : "compute_graph.Layer_18_wts_ddr",
261
+ "offset" : 5248
262
+ },
263
+ {
264
+ "name" : "compute_graph.Layer_19_wts_ddr",
265
+ "offset" : 7424
266
+ },
267
+ {
268
+ "name" : "compute_graph.Layer_20_wts_ddr",
269
+ "offset" : 10688
270
+ },
271
+ {
272
+ "name" : "compute_graph.Layer_22_wts_ddr",
273
+ "offset" : 11456
274
+ },
275
+ {
276
+ "name" : "compute_graph.Layer_23_wts_ddr",
277
+ "offset" : 13888
278
+ },
279
+ {
280
+ "name" : "compute_graph.Layer_24_wts_ddr",
281
+ "offset" : 17152
282
+ },
283
+ {
284
+ "name" : "compute_graph.Layer_27_wts_ddr",
285
+ "offset" : 19264
286
+ },
287
+ {
288
+ "name" : "compute_graph.Layer_28_wts_ddr",
289
+ "offset" : 21696
290
+ },
291
+ {
292
+ "name" : "compute_graph.Layer_34_wts_ddr",
293
+ "offset" : 26048
294
+ },
295
+ {
296
+ "name" : "compute_graph.Layer_35_wts_ddr",
297
+ "offset" : 28480
298
+ },
299
+ {
300
+ "name" : "compute_graph.Layer_36_wts_ddr",
301
+ "offset" : 32832
302
+ },
303
+ {
304
+ "name" : "compute_graph.Layer_39_wts_ddr",
305
+ "offset" : 35648
306
+ },
307
+ {
308
+ "name" : "compute_graph.Layer_40_wts_ddr",
309
+ "offset" : 39616
310
+ },
311
+ {
312
+ "name" : "compute_graph.Layer_46_wts_ddr",
313
+ "offset" : 43968
314
+ },
315
+ {
316
+ "name" : "compute_graph.Layer_47_wts_ddr",
317
+ "offset" : 48320
318
+ },
319
+ {
320
+ "name" : "compute_graph.Layer_48_wts_ddr",
321
+ "offset" : 52672
322
+ },
323
+ {
324
+ "name" : "compute_graph.Layer_51_wts_ddr",
325
+ "offset" : 55488
326
+ },
327
+ {
328
+ "name" : "compute_graph.Layer_52_wts_ddr",
329
+ "offset" : 59456
330
+ },
331
+ {
332
+ "name" : "compute_graph.Layer_58_wts_ddr",
333
+ "offset" : 63808
334
+ },
335
+ {
336
+ "name" : "compute_graph.Layer_59_wts_ddr",
337
+ "offset" : 68160
338
+ },
339
+ {
340
+ "name" : "compute_graph.Layer_65_wts_ddr",
341
+ "offset" : 76864
342
+ },
343
+ {
344
+ "name" : "compute_graph.Layer_70_wts_ddr",
345
+ "offset" : 78912
346
+ },
347
+ {
348
+ "name" : "compute_graph.Layer_71_wts_ddr",
349
+ "offset" : 91008
350
+ },
351
+ {
352
+ "name" : "compute_graph.Layer_76_wts_ddr",
353
+ "offset" : 101760
354
+ },
355
+ {
356
+ "name" : "compute_graph.Layer_81_wts_ddr",
357
+ "offset" : 103552
358
+ },
359
+ {
360
+ "name" : "compute_graph.Layer_82_wts_ddr",
361
+ "offset" : 113920
362
+ },
363
+ {
364
+ "name" : "compute_graph.Layer_87_wts_ddr",
365
+ "offset" : 121984
366
+ },
367
+ {
368
+ "name" : "compute_graph.Layer_92_wts_ddr",
369
+ "offset" : 123520
370
+ },
371
+ {
372
+ "name" : "compute_graph.Layer_93_wts_ddr",
373
+ "offset" : 133120
374
+ },
375
+ {
376
+ "name" : "compute_graph.Layer_98_wts_ddr",
377
+ "offset" : 141184
378
+ },
379
+ {
380
+ "name" : "compute_graph.Layer_103_wts_ddr",
381
+ "offset" : 142720
382
+ },
383
+ {
384
+ "name" : "compute_graph.Layer_104_wts_ddr",
385
+ "offset" : 152320
386
+ },
387
+ {
388
+ "name" : "compute_graph.Layer_109_wts_ddr",
389
+ "offset" : 172480
390
+ },
391
+ {
392
+ "name" : "compute_graph.Layer_116_wts_ddr",
393
+ "offset" : 176320
394
+ },
395
+ {
396
+ "name" : "compute_graph.Layer_117_wts_ddr",
397
+ "offset" : 208064
398
+ },
399
+ {
400
+ "name" : "compute_graph.Layer_123_wts_ddr",
401
+ "offset" : 239808
402
+ },
403
+ {
404
+ "name" : "compute_graph.Layer_124_wts_ddr",
405
+ "offset" : 271808
406
+ },
407
+ {
408
+ "name" : "compute_graph.Layer_129_wts_ddr",
409
+ "offset" : 317504
410
+ },
411
+ {
412
+ "name" : "compute_graph.Layer_136_wts_ddr",
413
+ "offset" : 322880
414
+ },
415
+ {
416
+ "name" : "compute_graph.Layer_137_wts_ddr",
417
+ "offset" : 390080
418
+ },
419
+ {
420
+ "name" : "compute_graph.Layer_143_wts_ddr",
421
+ "offset" : 450624
422
+ },
423
+ {
424
+ "name" : "compute_graph.Layer_144_wts_ddr",
425
+ "offset" : 495424
426
+ },
427
+ {
428
+ "name" : "compute_graph.Layer_149_wts_ddr",
429
+ "offset" : 541120
430
+ },
431
+ {
432
+ "name" : "compute_graph.Layer_156_wts_ddr",
433
+ "offset" : 578752
434
+ },
435
+ {
436
+ "name" : "compute_graph.Layer_157_wts_ddr",
437
+ "offset" : 645952
438
+ },
439
+ {
440
+ "name" : "compute_graph.Layer_163_wts_ddr",
441
+ "offset" : 706496
442
+ },
443
+ {
444
+ "name" : "compute_graph.Layer_164_wts_ddr",
445
+ "offset" : 766976
446
+ },
447
+ {
448
+ "name" : "compute_graph.Layer_169_wts_ddr",
449
+ "offset" : 847616
450
+ },
451
+ {
452
+ "name" : "compute_graph.Layer_176_wts_ddr",
453
+ "offset" : 901376
454
+ },
455
+ {
456
+ "name" : "compute_graph.Layer_177_wts_ddr",
457
+ "offset" : 1026816
458
+ },
459
+ {
460
+ "name" : "compute_graph.Layer_183_wts_ddr",
461
+ "offset" : 1147776
462
+ },
463
+ {
464
+ "name" : "compute_graph.Layer_184_wts_ddr",
465
+ "offset" : 1248000
466
+ },
467
+ {
468
+ "name" : "compute_graph.Layer_189_wts_ddr",
469
+ "offset" : 1328640
470
+ },
471
+ {
472
+ "name" : "compute_graph.Layer_196_wts_ddr",
473
+ "offset" : 1382400
474
+ },
475
+ {
476
+ "name" : "compute_graph.Layer_197_wts_ddr",
477
+ "offset" : 1507840
478
+ },
479
+ {
480
+ "name" : "compute_graph.Layer_203_wts_ddr",
481
+ "offset" : 1628800
482
+ },
483
+ {
484
+ "name" : "compute_graph.Layer_204_wts_ddr",
485
+ "offset" : 1729024
486
+ },
487
+ {
488
+ "name" : "compute_graph.Layer_211_wts_ddr",
489
+ "offset" : 1809664
490
+ },
491
+ {
492
+ "name" : "compute_graph.Layer_214_wts_ddr",
493
+ "offset" : 1872384
494
+ },
495
+ {
496
+ "name" : "compute_graph.Layer_218_wts_ddr",
497
+ "offset" : 1936384
498
+ },
499
+ {
500
+ "name" : "compute_graph.Layer_226_wts_ddr",
501
+ "offset" : 2012160
502
+ },
503
+ {
504
+ "name" : "compute_graph.Layer_237_wts_ddr",
505
+ "offset" : 2050048
506
+ },
507
+ {
508
+ "name" : "compute_graph.Layer_241_wts_ddr",
509
+ "offset" : 2154240
510
+ },
511
+ {
512
+ "name" : "compute_graph.Layer_249_wts_ddr",
513
+ "offset" : 2201600
514
+ },
515
+ {
516
+ "name" : "compute_graph.Layer_257_wts_ddr",
517
+ "offset" : 2225280
518
+ },
519
+ {
520
+ "name" : "compute_graph.Layer_261_wts_ddr",
521
+ "offset" : 2258432
522
+ },
523
+ {
524
+ "name" : "compute_graph.Layer_269_wts_ddr",
525
+ "offset" : 2272640
526
+ },
527
+ {
528
+ "name" : "compute_graph.Layer_276_wts_ddr",
529
+ "offset" : 2278720
530
+ },
531
+ {
532
+ "name" : "compute_graph.Layer_280_wts_ddr",
533
+ "offset" : 2288192
534
+ },
535
+ {
536
+ "name" : "compute_graph.Layer_288_wts_ddr",
537
+ "offset" : 2292928
538
+ },
539
+ {
540
+ "name" : "compute_graph.Layer_295_wts_ddr",
541
+ "offset" : 2297664
542
+ },
543
+ {
544
+ "name" : "compute_graph.Layer_296_wts_ddr",
545
+ "offset" : 2303744
546
+ },
547
+ {
548
+ "name" : "compute_graph.Layer_297_wts_ddr",
549
+ "offset" : 2306112
550
+ },
551
+ {
552
+ "name" : "compute_graph.const_ifm_ddr_5",
553
+ "offset" : 2307200
554
+ },
555
+ {
556
+ "name" : "compute_graph.const_ifm_ddr_4",
557
+ "offset" : 2537600
558
+ },
559
+ {
560
+ "name" : "compute_graph.const_ifm_ddr_3",
561
+ "offset" : 2768000
562
+ },
563
+ {
564
+ "name" : "compute_graph.const_ifm_ddr_2",
565
+ "offset" : 2775680
566
+ },
567
+ {
568
+ "name" : "compute_graph.const_ifm_ddr_1",
569
+ "offset" : 2794080
570
+ },
571
+ {
572
+ "name" : "compute_graph.const_ifm_ddr",
573
+ "offset" : 2837280
574
+ }
575
+ ]
576
+ },
577
+ "spills" : {
578
+ "layers" : [
579
+ {
580
+ "name" : "compute_graph.l2l3_1_spill",
581
+ "offset" : 0
582
+ },
583
+ {
584
+ "name" : "compute_graph.l2l3_2_spill",
585
+ "offset" : 117760
586
+ },
587
+ {
588
+ "name" : "compute_graph.l2l3_scratch_0_3_spill",
589
+ "offset" : 206080
590
+ },
591
+ {
592
+ "name" : "compute_graph.l2l3_3_spill",
593
+ "offset" : 441600
594
+ },
595
+ {
596
+ "name" : "compute_graph.l2l3_4_spill",
597
+ "offset" : 677120
598
+ },
599
+ {
600
+ "name" : "compute_graph.l2l3_scratch_0_5_spill",
601
+ "offset" : 912640
602
+ },
603
+ {
604
+ "name" : "compute_graph.l2l3_5_spill",
605
+ "offset" : 1143040
606
+ },
607
+ {
608
+ "name" : "compute_graph.l2l3_6_spill",
609
+ "offset" : 1373440
610
+ },
611
+ {
612
+ "name" : "compute_graph.l2l3_7_spill",
613
+ "offset" : 1603840
614
+ },
615
+ {
616
+ "name" : "compute_graph.l2l3_8_spill",
617
+ "offset" : 1834240
618
+ },
619
+ {
620
+ "name" : "compute_graph.l2l3_9_spill",
621
+ "offset" : 1949440
622
+ },
623
+ {
624
+ "name" : "compute_graph.l2l3_10_spill",
625
+ "offset" : 2064640
626
+ },
627
+ {
628
+ "name" : "compute_graph.l2l3_11_spill",
629
+ "offset" : 2179840
630
+ },
631
+ {
632
+ "name" : "compute_graph.l2l3_12_spill",
633
+ "offset" : 2295040
634
+ },
635
+ {
636
+ "name" : "compute_graph.l2l3_13_spill",
637
+ "offset" : 2410240
638
+ },
639
+ {
640
+ "name" : "compute_graph.l2l3_14_spill",
641
+ "offset" : 2525440
642
+ },
643
+ {
644
+ "name" : "compute_graph.l2l3_15_spill",
645
+ "offset" : 2640640
646
+ },
647
+ {
648
+ "name" : "compute_graph.l2l3_16_spill",
649
+ "offset" : 3101440
650
+ },
651
+ {
652
+ "name" : "compute_graph.l2l3_scratch_0_17_spill",
653
+ "offset" : 3224320
654
+ },
655
+ {
656
+ "name" : "compute_graph.l2l3_17_spill",
657
+ "offset" : 3339520
658
+ },
659
+ {
660
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_1818",
661
+ "offset" : 3454720
662
+ },
663
+ {
664
+ "name" : "compute_graph.l2l3_19_spill",
665
+ "offset" : 3497920
666
+ },
667
+ {
668
+ "name" : "compute_graph.l2l3_20_spill",
669
+ "offset" : 3627520
670
+ },
671
+ {
672
+ "name" : "compute_graph.l2l3_scratch_0_21_spill",
673
+ "offset" : 3760000
674
+ },
675
+ {
676
+ "name" : "compute_graph.l2l3_21_spill",
677
+ "offset" : 3889600
678
+ },
679
+ {
680
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_2424",
681
+ "offset" : 4019200
682
+ },
683
+ {
684
+ "name" : "compute_graph.l2l3_24_spill",
685
+ "offset" : 4052320
686
+ },
687
+ {
688
+ "name" : "compute_graph.l2l3_25_spill",
689
+ "offset" : 4085440
690
+ },
691
+ {
692
+ "name" : "compute_graph.l2l3_31_spill",
693
+ "offset" : 4118560
694
+ },
695
+ {
696
+ "name" : "compute_graph.l2l3_scratch_0_32_spill",
697
+ "offset" : 4118596
698
+ },
699
+ {
700
+ "name" : "compute_graph.l2l3_scratch_1_32_spill",
701
+ "offset" : 4151716
702
+ },
703
+ {
704
+ "name" : "compute_graph.l2l3_32_spill",
705
+ "offset" : 4184836
706
+ },
707
+ {
708
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_3434",
709
+ "offset" : 4217956
710
+ },
711
+ {
712
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_3636",
713
+ "offset" : 4236356
714
+ },
715
+ {
716
+ "name" : "compute_graph.l2l3_36_spill",
717
+ "offset" : 4291556
718
+ },
719
+ {
720
+ "name" : "compute_graph.l2l3_37_spill",
721
+ "offset" : 4346756
722
+ },
723
+ {
724
+ "name" : "compute_graph.l2l3_43_spill",
725
+ "offset" : 4401956
726
+ },
727
+ {
728
+ "name" : "compute_graph.l2l3_scratch_0_44_spill",
729
+ "offset" : 4402016
730
+ },
731
+ {
732
+ "name" : "compute_graph.l2l3_scratch_1_44_spill",
733
+ "offset" : 4457216
734
+ },
735
+ {
736
+ "name" : "compute_graph.l2l3_44_spill",
737
+ "offset" : 4512416
738
+ },
739
+ {
740
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_4646",
741
+ "offset" : 4567616
742
+ },
743
+ {
744
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_4848",
745
+ "offset" : 4586016
746
+ },
747
+ {
748
+ "name" : "compute_graph.l2l3_48_spill",
749
+ "offset" : 4641216
750
+ },
751
+ {
752
+ "name" : "compute_graph.l2l3_49_spill",
753
+ "offset" : 4696416
754
+ },
755
+ {
756
+ "name" : "compute_graph.l2l3_55_spill",
757
+ "offset" : 4751616
758
+ },
759
+ {
760
+ "name" : "compute_graph.l2l3_scratch_0_56_spill",
761
+ "offset" : 4751676
762
+ },
763
+ {
764
+ "name" : "compute_graph.l2l3_scratch_1_56_spill",
765
+ "offset" : 4806876
766
+ },
767
+ {
768
+ "name" : "compute_graph.l2l3_56_spill",
769
+ "offset" : 4862076
770
+ },
771
+ {
772
+ "name" : "compute_graph.l2l3_59_spill",
773
+ "offset" : 4917276
774
+ },
775
+ {
776
+ "name" : "compute_graph.l2l3_62_spill",
777
+ "offset" : 5027676
778
+ },
779
+ {
780
+ "name" : "compute_graph.l2l3_63_spill",
781
+ "offset" : 5138076
782
+ },
783
+ {
784
+ "name" : "compute_graph.l2l3_scratch_0_64_spill",
785
+ "offset" : 5253276
786
+ },
787
+ {
788
+ "name" : "compute_graph.l2l3_64_spill",
789
+ "offset" : 5363676
790
+ },
791
+ {
792
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_113113",
793
+ "offset" : 5474076
794
+ },
795
+ {
796
+ "name" : "compute_graph.l2l3_113_spill",
797
+ "offset" : 5531676
798
+ },
799
+ {
800
+ "name" : "compute_graph.l2l3_114_spill",
801
+ "offset" : 5589276
802
+ },
803
+ {
804
+ "name" : "compute_graph.l2l3_120_spill",
805
+ "offset" : 5646876
806
+ },
807
+ {
808
+ "name" : "compute_graph.l2l3_scratch_0_121_spill",
809
+ "offset" : 5647116
810
+ },
811
+ {
812
+ "name" : "compute_graph.l2l3_scratch_1_121_spill",
813
+ "offset" : 5704716
814
+ },
815
+ {
816
+ "name" : "compute_graph.l2l3_121_spill",
817
+ "offset" : 5762316
818
+ },
819
+ {
820
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_123123",
821
+ "offset" : 5819916
822
+ },
823
+ {
824
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_133133",
825
+ "offset" : 5833356
826
+ },
827
+ {
828
+ "name" : "compute_graph.l2l3_133_spill",
829
+ "offset" : 5913996
830
+ },
831
+ {
832
+ "name" : "compute_graph.l2l3_134_spill",
833
+ "offset" : 5994636
834
+ },
835
+ {
836
+ "name" : "compute_graph.l2l3_140_spill",
837
+ "offset" : 6075276
838
+ },
839
+ {
840
+ "name" : "compute_graph.l2l3_scratch_0_141_spill",
841
+ "offset" : 6075612
842
+ },
843
+ {
844
+ "name" : "compute_graph.l2l3_scratch_1_141_spill",
845
+ "offset" : 6156252
846
+ },
847
+ {
848
+ "name" : "compute_graph.l2l3_141_spill",
849
+ "offset" : 6236892
850
+ },
851
+ {
852
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_153153",
853
+ "offset" : 6317532
854
+ },
855
+ {
856
+ "name" : "compute_graph.l2l3_153_spill",
857
+ "offset" : 6398172
858
+ },
859
+ {
860
+ "name" : "compute_graph.l2l3_154_spill",
861
+ "offset" : 6478812
862
+ },
863
+ {
864
+ "name" : "compute_graph.l2l3_160_spill",
865
+ "offset" : 6559452
866
+ },
867
+ {
868
+ "name" : "compute_graph.l2l3_scratch_1_161_spill",
869
+ "offset" : 6559788
870
+ },
871
+ {
872
+ "name" : "compute_graph.l2l3_scratch_0_161_spill",
873
+ "offset" : 6640428
874
+ },
875
+ {
876
+ "name" : "compute_graph.l2l3_161_spill",
877
+ "offset" : 6721068
878
+ },
879
+ {
880
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_163163",
881
+ "offset" : 6801708
882
+ },
883
+ {
884
+ "name" : "compute_graph.l2l3_164_spill",
885
+ "offset" : 6820908
886
+ },
887
+ {
888
+ "name" : "compute_graph.l2l3_167_spill",
889
+ "offset" : 6936108
890
+ },
891
+ {
892
+ "name" : "compute_graph.l2l3_168_spill",
893
+ "offset" : 7051308
894
+ },
895
+ {
896
+ "name" : "compute_graph.l2l3_169_spill",
897
+ "offset" : 7166508
898
+ },
899
+ {
900
+ "name" : "compute_graph.l2l3_172_spill",
901
+ "offset" : 7281708
902
+ },
903
+ {
904
+ "name" : "compute_graph.l2l3_173_spill",
905
+ "offset" : 7396908
906
+ },
907
+ {
908
+ "name" : "compute_graph.l2l3_174_spill",
909
+ "offset" : 7512108
910
+ },
911
+ {
912
+ "name" : "compute_graph.l2l3_180_spill",
913
+ "offset" : 7627308
914
+ },
915
+ {
916
+ "name" : "compute_graph.l2l3_scratch_0_181_spill",
917
+ "offset" : 7627788
918
+ },
919
+ {
920
+ "name" : "compute_graph.l2l3_scratch_1_181_spill",
921
+ "offset" : 7742988
922
+ },
923
+ {
924
+ "name" : "compute_graph.l2l3_181_spill",
925
+ "offset" : 7858188
926
+ },
927
+ {
928
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_183183",
929
+ "offset" : 7973388
930
+ },
931
+ {
932
+ "name" : "compute_graph.l2l3_184_spill",
933
+ "offset" : 7992588
934
+ },
935
+ {
936
+ "name" : "compute_graph.l2l3_187_spill",
937
+ "offset" : 8107788
938
+ },
939
+ {
940
+ "name" : "compute_graph.l2l3_188_spill",
941
+ "offset" : 8222988
942
+ },
943
+ {
944
+ "name" : "compute_graph.l2l3_189_spill",
945
+ "offset" : 8338188
946
+ },
947
+ {
948
+ "name" : "compute_graph.l2l3_192_spill",
949
+ "offset" : 8453388
950
+ },
951
+ {
952
+ "name" : "compute_graph.l2l3_193_spill",
953
+ "offset" : 8568588
954
+ },
955
+ {
956
+ "name" : "compute_graph.l2l3_194_spill",
957
+ "offset" : 8683788
958
+ },
959
+ {
960
+ "name" : "compute_graph.l2l3_200_spill",
961
+ "offset" : 8798988
962
+ },
963
+ {
964
+ "name" : "compute_graph.l2l3_scratch_0_201_spill",
965
+ "offset" : 8799468
966
+ },
967
+ {
968
+ "name" : "compute_graph.l2l3_scratch_1_201_spill",
969
+ "offset" : 8914668
970
+ },
971
+ {
972
+ "name" : "compute_graph.l2l3_201_spill",
973
+ "offset" : 9029868
974
+ },
975
+ {
976
+ "name" : "compute_graph.l2l3_204_spill",
977
+ "offset" : 9145068
978
+ },
979
+ {
980
+ "name" : "compute_graph.l2l3_207_spill",
981
+ "offset" : 9260268
982
+ },
983
+ {
984
+ "name" : "compute_graph.l2l3_208_spill",
985
+ "offset" : 9375468
986
+ },
987
+ {
988
+ "name" : "compute_graph.l2l3_209_spill",
989
+ "offset" : 9490668
990
+ },
991
+ {
992
+ "name" : "compute_graph.l2l3_212_spill",
993
+ "offset" : 9605868
994
+ },
995
+ {
996
+ "name" : "compute_graph.l2l3_scratch_0_213_spill",
997
+ "offset" : 9605932
998
+ },
999
+ {
1000
+ "name" : "compute_graph.l2l3_scratch_1_213_spill",
1001
+ "offset" : 9621292
1002
+ },
1003
+ {
1004
+ "name" : "compute_graph.l2l3_213_spill",
1005
+ "offset" : 9636652
1006
+ },
1007
+ {
1008
+ "name" : "compute_graph.l2l3_214_spill",
1009
+ "offset" : 9652012
1010
+ },
1011
+ {
1012
+ "name" : "compute_graph.l2l3_215_spill",
1013
+ "offset" : 9667372
1014
+ },
1015
+ {
1016
+ "name" : "compute_graph.l2l3_216_spill",
1017
+ "offset" : 9675052
1018
+ },
1019
+ {
1020
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_217",
1021
+ "offset" : 9682732
1022
+ },
1023
+ {
1024
+ "name" : "compute_graph.l2l3_219_spill",
1025
+ "offset" : 9698092
1026
+ },
1027
+ {
1028
+ "name" : "compute_graph.l2l3_220_spill",
1029
+ "offset" : 9713452
1030
+ },
1031
+ {
1032
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_222222",
1033
+ "offset" : 9721132
1034
+ },
1035
+ {
1036
+ "name" : "compute_graph.l2l3_223_spill",
1037
+ "offset" : 9728812
1038
+ },
1039
+ {
1040
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_225",
1041
+ "offset" : 9736492
1042
+ },
1043
+ {
1044
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_230",
1045
+ "offset" : 9751852
1046
+ },
1047
+ {
1048
+ "name" : "compute_graph.l2l3_231_spill",
1049
+ "offset" : 9767212
1050
+ },
1051
+ {
1052
+ "name" : "compute_graph.l2l3_232_spill",
1053
+ "offset" : 9828652
1054
+ },
1055
+ {
1056
+ "name" : "compute_graph.l2l3_233_spill",
1057
+ "offset" : 9887532
1058
+ },
1059
+ {
1060
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_236",
1061
+ "offset" : 9945132
1062
+ },
1063
+ {
1064
+ "name" : "compute_graph.l2l3_237_spill",
1065
+ "offset" : 10026092
1066
+ },
1067
+ {
1068
+ "name" : "compute_graph.l2l3_238_spill",
1069
+ "offset" : 10062892
1070
+ },
1071
+ {
1072
+ "name" : "compute_graph.l2l3_239_spill",
1073
+ "offset" : 10081292
1074
+ },
1075
+ {
1076
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_240",
1077
+ "offset" : 10099692
1078
+ },
1079
+ {
1080
+ "name" : "compute_graph.l2l3_242_spill",
1081
+ "offset" : 10136492
1082
+ },
1083
+ {
1084
+ "name" : "compute_graph.l2l3_243_spill",
1085
+ "offset" : 10173292
1086
+ },
1087
+ {
1088
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_245245",
1089
+ "offset" : 10191692
1090
+ },
1091
+ {
1092
+ "name" : "compute_graph.l2l3_246_spill",
1093
+ "offset" : 10210092
1094
+ },
1095
+ {
1096
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_248",
1097
+ "offset" : 10228492
1098
+ },
1099
+ {
1100
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_253",
1101
+ "offset" : 10265292
1102
+ },
1103
+ {
1104
+ "name" : "compute_graph.l2l3_254_spill",
1105
+ "offset" : 10302092
1106
+ },
1107
+ {
1108
+ "name" : "compute_graph.l2l3_255_spill",
1109
+ "offset" : 10449292
1110
+ },
1111
+ {
1112
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_256",
1113
+ "offset" : 10593292
1114
+ },
1115
+ {
1116
+ "name" : "compute_graph.l2l3_257_spill",
1117
+ "offset" : 10794892
1118
+ },
1119
+ {
1120
+ "name" : "compute_graph.l2l3_258_spill",
1121
+ "offset" : 10866892
1122
+ },
1123
+ {
1124
+ "name" : "compute_graph.l2l3_259_spill",
1125
+ "offset" : 10910092
1126
+ },
1127
+ {
1128
+ "name" : "compute_graph.l2l3_260_spill",
1129
+ "offset" : 10953292
1130
+ },
1131
+ {
1132
+ "name" : "compute_graph.l2l3_262_spill",
1133
+ "offset" : 11025292
1134
+ },
1135
+ {
1136
+ "name" : "compute_graph.l2l3_263_spill",
1137
+ "offset" : 11097292
1138
+ },
1139
+ {
1140
+ "name" : "compute_graph.L3_OFM_Buffer_layer_TGSpilling_265265",
1141
+ "offset" : 11140492
1142
+ },
1143
+ {
1144
+ "name" : "compute_graph.l2l3_266_spill",
1145
+ "offset" : 11183692
1146
+ },
1147
+ {
1148
+ "name" : "compute_graph.l2l3_267_spill",
1149
+ "offset" : 11226892
1150
+ },
1151
+ {
1152
+ "name" : "compute_graph.l2l3_268_spill",
1153
+ "offset" : 11270092
1154
+ },
1155
+ {
1156
+ "name" : "compute_graph.l2l3_273_spill",
1157
+ "offset" : 11342092
1158
+ },
1159
+ {
1160
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_275",
1161
+ "offset" : 11414092
1162
+ },
1163
+ {
1164
+ "name" : "compute_graph.l2l3_276_spill",
1165
+ "offset" : 11874892
1166
+ },
1167
+ {
1168
+ "name" : "compute_graph.l2l3_277_spill",
1169
+ "offset" : 12105292
1170
+ },
1171
+ {
1172
+ "name" : "compute_graph.l2l3_278_spill",
1173
+ "offset" : 12220492
1174
+ },
1175
+ {
1176
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_279",
1177
+ "offset" : 12335692
1178
+ },
1179
+ {
1180
+ "name" : "compute_graph.l2l3_280_spill",
1181
+ "offset" : 12566092
1182
+ },
1183
+ {
1184
+ "name" : "compute_graph.l2l3_281_spill",
1185
+ "offset" : 12796492
1186
+ },
1187
+ {
1188
+ "name" : "compute_graph.l2l3_282_spill",
1189
+ "offset" : 13026892
1190
+ },
1191
+ {
1192
+ "name" : "compute_graph.l2l3_283_spill",
1193
+ "offset" : 13142092
1194
+ },
1195
+ {
1196
+ "name" : "compute_graph.l2l3_284_spill",
1197
+ "offset" : 13257292
1198
+ },
1199
+ {
1200
+ "name" : "compute_graph.l2l3_285_spill",
1201
+ "offset" : 13372492
1202
+ },
1203
+ {
1204
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_287",
1205
+ "offset" : 13487692
1206
+ },
1207
+ {
1208
+ "name" : "compute_graph.l2l3_288_spill",
1209
+ "offset" : 13718092
1210
+ },
1211
+ {
1212
+ "name" : "compute_graph.l2l3_289_spill",
1213
+ "offset" : 13833292
1214
+ },
1215
+ {
1216
+ "name" : "compute_graph.l2l3_290_spill",
1217
+ "offset" : 13948492
1218
+ },
1219
+ {
1220
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_292",
1221
+ "offset" : 14063692
1222
+ },
1223
+ {
1224
+ "name" : "compute_graph.spill_L3_Concat_Buffer_layer_294",
1225
+ "offset" : 14294092
1226
+ },
1227
+ {
1228
+ "name" : "compute_graph.l2l3_295_spill",
1229
+ "offset" : 15446092
1230
+ },
1231
+ {
1232
+ "name" : "compute_graph.l2l3_296_spill",
1233
+ "offset" : 15906892
1234
+ },
1235
+ {
1236
+ "name" : "compute_graph.l2l3_297_spill",
1237
+ "offset" : 16367692
1238
+ },
1239
+ {
1240
+ "name" : "compute_graph.l2l3_298_spill",
1241
+ "offset" : 16598092
1242
+ },
1243
+ {
1244
+ "name" : "compute_graph.l2l3_300_spill",
1245
+ "offset" : 16828492
1246
+ },
1247
+ {
1248
+ "name" : "compute_graph.l2l3_301_spill",
1249
+ "offset" : 17058892
1250
+ }
1251
+ ]
1252
+ }
1253
+ }
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/ml_txn.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4e8c431abc3fb04c2333b756f029894d8440645eeb714fb234525160657d575e
3
+ size 744144
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/model_top.cpp ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ #include "multi_layer_overlay.h"
2
+ FlexMLGraph compute_graph;
3
+ #if defined(__AIESIM__) || defined(__X86SIM__)
4
+ #include "../aie_runtime_control.cpp"
5
+ int main() { return 0; }
6
+ #endif
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/par.mlopslib.tosa.mlir ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/par.subgraph.dse.mlir ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/par.subgraph.pre-dse.mlir ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/tensor_dims.json ADDED
The diff for this file is too large to render. See raw diff
 
segmentation_1_4_0_fp32_combined/vaiml_par_0/0/unified-4x4.xclbin ADDED
Binary file (47.5 kB). View file