1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
| import os
import shutil
import subprocess
import zipfile
import torch
import torch.nn as nn
from torch.library import triton_op, wrap_triton
import triton
import triton.language as tl
@triton.jit
def _triton_copy_kernel(
in_ptr,
out_ptr,
n_elements,
BLOCK_SIZE: "tl.constexpr",
):
pid = tl.program_id(axis=0)
block_start = pid * BLOCK_SIZE
offsets = block_start + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
x = tl.load(in_ptr + offsets, mask=mask)
tl.store(out_ptr + offsets, x, mask=mask)
@triton_op("custom_ops::triton_copy", mutates_args=())
def triton_copy_op(x: torch.Tensor) -> torch.Tensor:
out = torch.empty_like(x)
if x.is_cuda:
n_elements = x.numel()
def grid(meta):
return (triton.cdiv(n_elements, meta["BLOCK_SIZE"]), )
wrap_triton(_triton_copy_kernel)[grid](x,
out,
n_elements,
BLOCK_SIZE=1024)
else:
out.copy_(x)
return out
class VectorCopyModel(nn.Module):
def forward(self, x: torch.Tensor) -> torch.Tensor:
return triton_copy_op(x)
def main():
device = "cuda" if torch.cuda.is_available() else "cpu"
if device != "cuda":
raise RuntimeError(
"CUDA device required for AOTInductor Triton compilation.")
print("\n" + "=" * 75)
print(
" PIPELINE: Export -> AOTI Compile -> Bitwise Test -> SASS Inspection")
print("=" * 75)
model = VectorCopyModel().to(device).eval()
example_args = (torch.randn(1024 * 1024,
device=device,
dtype=torch.float32), )
print(
"\n[STEP 1] Exporting graph with torch.export.export(..., strict=True)"
)
ep = torch.export.export(model, args=example_args, strict=True)
decomposed_ep = ep.run_decompositions()
print(" ✅ ExportedProgram graph successfully captured and decomposed.")
print("\n[STEP 2] Compiling to AOTI Package (.pt2 artifact)")
output_dir = os.path.abspath("./aoti_output")
os.makedirs(output_dir, exist_ok=True)
pkg_path = os.path.join(output_dir, "model.pt2")
compiled_pkg = torch._inductor.aoti_compile_and_package(
decomposed_ep, package_path=pkg_path)
print(f" ✅ Compiled Package Generated: {compiled_pkg}")
print("\n[STEP 3] Executing Compiled Model and Verifying Bitwise Identity")
runner = torch._inductor.aoti_load_package(compiled_pkg)
test_input = torch.randn(1024 * 1024, device=device, dtype=torch.float32)
output = runner(test_input)
assert torch.equal(test_input, output), "FAIL: Tensors are not equal!"
input_bits = test_input.view(torch.int32)
output_bits = output.view(torch.int32)
mismatches = (input_bits != output_bits).sum().item()
assert mismatches == 0, f"FAIL: Found {mismatches} bitwise mismatch(es)!"
print(
" ✅ Bitwise Identity Confirmed: Input and Output are 100% bitwise identical!"
)
print("\n[STEP 4] Unpacking .pt2 Archive and Scanning Artifacts")
extract_dir = os.path.join(output_dir, "extracted")
os.makedirs(extract_dir, exist_ok=True)
extracted_cubins = []
extracted_so = None
extracted_cpp = None
with zipfile.ZipFile(compiled_pkg, "r") as zip_ref:
zip_ref.extractall(extract_dir)
for root, _, files in os.walk(extract_dir):
for f in files:
full_path = os.path.join(root, f)
if f.endswith(".cubin"):
extracted_cubins.append(full_path)
elif f.endswith(".wrapper.so") or (f.endswith(".so")
and not extracted_so):
extracted_so = full_path
elif f.endswith(".wrapper.cpp") or (f.endswith(".cpp")
and not extracted_cpp):
extracted_cpp = full_path
print(f" ✅ Found {len(extracted_cubins)} .cubin file(s) in package.")
if extracted_cpp:
print(f" ✅ Found wrapper C++: {extracted_cpp}")
if extracted_so:
print(f" ✅ Found wrapper .so: {extracted_so}")
print("\n[STEP 5] Running SASS Disassembly (nvdisasm) on Extracted .cubin")
if extracted_cubins and shutil.which("nvdisasm"):
for cubin_path in extracted_cubins:
cubin_name = os.path.basename(cubin_path)
res = subprocess.run(
["nvdisasm", "-g", cubin_path],
capture_output=True,
text=True,
)
sass_output = res.stdout
if "LDG.E.128" in sass_output or "STG.E.128" in sass_output:
print(
f" ✅ [nvdisasm CONFIRMED in {cubin_name}] Found 128-bit vector instructions:"
)
for line in sass_output.splitlines():
if "LDG.E.128" in line or "STG.E.128" in line:
print(f" {line.strip()}")
else:
print(f" --> Disassembly completed for {cubin_name}.")
elif not shutil.which("nvdisasm"):
print(" --> [SKIPPED] 'nvdisasm' not found in system PATH.")
print(
"\n[STEP 6] Running SASS Disassembly (cuobjdump) on Extracted .cubin")
if extracted_cubins and shutil.which("cuobjdump"):
for cubin_path in extracted_cubins:
cubin_name = os.path.basename(cubin_path)
res = subprocess.run(
["cuobjdump", "-sass", cubin_path],
capture_output=True,
text=True,
)
cubin_sass = res.stdout
if "LDG.E.128" in cubin_sass or "STG.E.128" in cubin_sass:
print(
f" ✅ [cuobjdump CONFIRMED in {cubin_name}] Found 128-bit vector instructions:"
)
for line in cubin_sass.splitlines():
if "LDG.E.128" in line or "STG.E.128" in line:
print(f" {line.strip()}")
else:
print(f" --> Disassembly of {cubin_name} completed.")
elif not shutil.which("cuobjdump"):
print(" --> [SKIPPED] 'cuobjdump' not found in system PATH.")
if __name__ == "__main__":
main()
|