Repository navigation
Expand file tree
/
Copy pathconvolution.py
More file actions
377 lines (358 loc) · 17.1 KB
/
Copy pathconvolution.py
File metadata and controls
377 lines (358 loc) · 17.1 KB
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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
"""Conservative single-convolution reference lowering;fresh verification required.
Forward FP32 only,on Ampere-or-newer CUDA. The generated module initializes
ordinary Parameters with PyTorch's initialization equations;all convolution
arithmetic runs in Triton. No reference code is executed during analysis.
"""
import ast, copy
from infinicoder.domains.operator.pointwise import Proposal
def propose(task):
if "REFERENCE:\n" not in task or not ("FP32" in task or "float32" in task):
return None
instructions, source = task.split("REFERENCE:\n", 1)
if any(
word in instructions.lower()
for word in ["backward", "autograd", "gradient", "training", "训练"]
):
return None
if len(source) > 65536:
return None
try:
tree = ast.parse(source)
if sum(1 for _ in ast.walk(tree)) > 3000:
return None
# Resolve only explicit official torch imports;reject shadowed aliases.
modules = {}
for n in tree.body:
if isinstance(n, ast.Import):
for a in n.names:
modules[a.asname or a.name] = a.name
elif isinstance(n, ast.ImportFrom) and n.module == "torch":
for a in n.names:
if a.name == "nn":
modules[a.asname or "nn"] = "torch.nn"
def qualified(node):
if isinstance(node, ast.Name):
return modules.get(node.id, node.id)
if isinstance(node, ast.Attribute):
return qualified(node.value) + "." + node.attr
return ""
if not any(value == "torch" for value in modules.values()):
return None
if any(
key in modules and modules[key] != value
for key, value in [("torch", "torch"), ("nn", "torch.nn"), ("math", "math")]
):
return None
for n in ast.walk(tree):
if isinstance(n, ast.Name) and n.id.startswith("_ic_"):
return None
if (
isinstance(n, ast.Name)
and isinstance(n.ctx, (ast.Store, ast.Del))
and n.id in modules
):
return None
if (
isinstance(n, ast.Attribute)
and isinstance(n.ctx, (ast.Store, ast.Del))
and qualified(n.value).split(".")[0] == "torch"
):
return None
cls = next(n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "Model")
if (
cls.decorator_list
or len(cls.bases) != 1
or qualified(cls.bases[0]) != "torch.nn.Module"
):
return None
methods = {n.name: n for n in cls.body if isinstance(n, ast.FunctionDef)}
if set(methods) != {"__init__", "forward"}:
return None
forward = methods["forward"]
args = forward.args
if (
len(args.args) != 2
or args.posonlyargs
or args.vararg
or args.kwarg
or args.kwonlyargs
or args.defaults
or forward.decorator_list
):
return None
body = [
n
for n in forward.body
if not (
isinstance(n, ast.Expr)
and isinstance(n.value, ast.Constant)
and isinstance(n.value.value, str)
)
]
if len(body) != 1 or not isinstance(body[0], ast.Return):
return None
call = body[0].value
if (
not isinstance(call, ast.Call)
or call.keywords
or len(call.args) != 1
or not isinstance(call.args[0], ast.Name)
or call.args[0].id != args.args[1].arg
):
return None
if (
not isinstance(call.func, ast.Attribute)
or not isinstance(call.func.value, ast.Name)
or call.func.value.id != args.args[0].arg
):
return None
field = call.func.attr
constructors = []
for n in methods["__init__"].body:
if (
isinstance(n, ast.Assign)
and len(n.targets) == 1
and isinstance(n.targets[0], ast.Attribute)
and ast.unparse(n.targets[0]) == "self." + field
and isinstance(n.value, ast.Call)
):
constructors.append(n.value)
if len(constructors) != 1:
return None
constructor = constructors[0]
name = qualified(constructor.func).removeprefix("torch.nn.")
if (
name
not in [
"Conv1d",
"Conv2d",
"Conv3d",
"ConvTranspose1d",
"ConvTranspose2d",
"ConvTranspose3d",
]
or qualified(constructor.func) != "torch.nn." + name
):
return None
# The initializer must only initialize its superclass and this one layer.
statements = [
n
for n in methods["__init__"].body
if not (
isinstance(n, ast.Expr)
and isinstance(n.value, ast.Constant)
and isinstance(n.value.value, str)
)
]
for n in statements:
if isinstance(n, ast.Assign) and n.value is constructor:
continue
if (
isinstance(n, ast.Expr)
and isinstance(n.value, ast.Call)
and ast.unparse(n.value.func) in ["super().__init__", "super(Model, self).__init__"]
and not n.value.args
and not n.value.keywords
):
continue
return None
for keyword in constructor.keywords:
if keyword.arg not in [
"in_channels",
"out_channels",
"kernel_size",
"stride",
"padding",
"dilation",
"groups",
"bias",
"output_padding",
"padding_mode",
"device",
"dtype",
]:
return None
if (
keyword.arg == "padding"
and isinstance(keyword.value, ast.Constant)
and isinstance(keyword.value.value, str)
):
return None
rank = int(name[-2])
transposed = "Transpose" in name
new = copy.deepcopy(cls)
class Rewrite(ast.NodeTransformer):
def visit_Name(self, node):
if node.id == "Model":
node.id = "ModelNew"
return node
def visit_Call(self, node):
node = self.generic_visit(node)
if qualified(node.func) == "torch.nn." + name:
node.func = ast.Name("_ic_" + name, ast.Load())
return node
def visit_Expr(self, node):
if isinstance(node.value, ast.Constant) and isinstance(node.value.value, str):
return None
return self.generic_visit(node)
new = Rewrite().visit(new)
new.name = "ModelNew"
f = next(n for n in new.body if isinstance(n, ast.FunctionDef) and n.name == "forward")
f.body = ast.parse(
"return _ic_forward(" + args.args[1].arg + ", " + args.args[0].arg + "." + field + ")"
).body
imports = "\n".join(
ast.unparse(n)
for n in tree.body
if (
isinstance(n, ast.Import)
and all(a.name in ["torch", "torch.nn", "math", "typing"] for a in n.names)
)
or (isinstance(n, ast.ImportFrom) and n.module in ["torch", "typing", "math"])
)
imports += "\nimport torch"
if "torch.nn as nn" not in imports:
imports += "\nimport torch.nn as nn"
factory = _factory(name, rank, transposed)
result = (
imports
+ "\n"
+ _RUNTIME
+ "\n"
+ factory
+ "\n"
+ ast.unparse(ast.fix_missing_locations(new))
+ "\n"
)
return Proposal(
result,
"single " + name,
(
"single stateful convolution;exact parameter initialization",
"FP32 forward only;no autograd",
"CUDA compute capability>=8",
"numeric nonnegative padding;positive spatial dimensions",
"independent input/weight strides;groups,dilation,transpose output padding",
),
provenance="static_single_convolution_lowering_v1",
)
except (
SyntaxError,
ValueError,
TypeError,
KeyError,
StopIteration,
AttributeError,
IndexError,
):
return None
def _factory(name, rank, transposed):
if transposed:
signature = 'in_channels,out_channels,kernel_size,stride=1,padding=0,output_padding=0,groups=1,bias=True,dilation=1,padding_mode="zeros",device=None,dtype=None'
else:
signature = 'in_channels,out_channels,kernel_size,stride=1,padding=0,dilation=1,groups=1,bias=True,padding_mode="zeros",device=None,dtype=None'
return (
"def _ic_"
+ name
+ "("
+ signature
+ "):\n return _ic_parameters(in_channels,out_channels,kernel_size,stride,padding,dilation,groups,bias,padding_mode,device,dtype,"
+ str(rank)
+ ","
+ str(transposed)
+ ","
+ ("output_padding" if transposed else "0")
+ ")\n"
)
_RUNTIME = """
import math
import triton
import triton.language as tl
def _ic_tuple(value,rank):
values=(value,)*rank if isinstance(value,int) else tuple(value)
if len(values)!=rank or any(type(v) is not int for v in values):raise ValueError('numeric convolution metadata required')
return values
def _ic_parameters(ci,co,kernel,stride,padding,dilation,groups,bias,padding_mode,device,dtype,rank,transposed,output_padding):
if padding_mode!='zeros' or groups<=0 or ci<=0 or co<=0 or ci%groups or co%groups:raise ValueError('unsupported convolution contract')
k=_ic_tuple(kernel,rank);s=_ic_tuple(stride,rank);p=_ic_tuple(padding,rank);d=_ic_tuple(dilation,rank);op=_ic_tuple(output_padding,rank)
if any(v<=0 for v in k+s+d) or any(v<0 for v in p+op):raise ValueError('invalid convolution metadata')
if transposed and any(o>=st and o>=di for o,st,di in zip(op,s,d)):raise ValueError('invalid output padding')
module=nn.Module();module.in_channels=ci;module.out_channels=co;module.groups=groups;module.kernel_size=k;module.stride=s;module.padding=p;module.dilation=d;module.output_padding=op;module.transposed=transposed
shape=(ci,co//groups,*k) if transposed else (co,ci//groups,*k)
module.weight=nn.Parameter(torch.empty(shape,device=device,dtype=dtype))
nn.init.kaiming_uniform_(module.weight,a=math.sqrt(5))
if bias:
module.bias=nn.Parameter(torch.empty(co,device=device,dtype=dtype))
bound=1/math.sqrt(shape[1]*math.prod(k));nn.init.uniform_(module.bias,-bound,bound)
else:module.register_parameter('bias',None)
return module
@triton.jit
def _ic_conv(X,W,Bias,Y,B:tl.constexpr,CI:tl.constexpr,CO:tl.constexpr,CG:tl.constexpr,OG:tl.constexpr,
ID:tl.constexpr,IH:tl.constexpr,IW:tl.constexpr,OD:tl.constexpr,OH:tl.constexpr,OW:tl.constexpr,
KD:tl.constexpr,KH:tl.constexpr,KW:tl.constexpr,SD:tl.constexpr,SH:tl.constexpr,SW:tl.constexpr,
PD:tl.constexpr,PH:tl.constexpr,PW:tl.constexpr,DD:tl.constexpr,DH:tl.constexpr,DW:tl.constexpr,
X0:tl.constexpr,X1:tl.constexpr,X2:tl.constexpr,X3:tl.constexpr,X4:tl.constexpr,
W0:tl.constexpr,W1:tl.constexpr,W2:tl.constexpr,W3:tl.constexpr,W4:tl.constexpr,
B0:tl.constexpr,HAS_BIAS:tl.constexpr,TRANSPOSED:tl.constexpr,BM:tl.constexpr,BN:tl.constexpr,BK:tl.constexpr):
m=(tl.program_id(0)*BM+tl.arange(0,BM)).to(tl.int64);n=tl.program_id(1)*BN+tl.arange(0,BN);group=tl.program_id(2)
volume=OD*OH*OW;batch=m//volume;spatial=m%volume;oz=spatial//(OH*OW);oy=spatial//OW%OH;ox=spatial%OW
acc=tl.full((BM,BN),0,tl.float32)
for part in range(tl.cdiv(CG*KD*KH*KW,BK)):
r=part*BK+tl.arange(0,BK);ic=r//(KD*KH*KW);z=r//(KH*KW)%KD;y=r//KW%KH;x=r%KW
if TRANSPOSED:
iz=oz[:,None]+PD-z[None,:]*DD;iy=oy[:,None]+PH-y[None,:]*DH;ix=ox[:,None]+PW-x[None,:]*DW
aligned=(iz%SD==0)&(iy%SH==0)&(ix%SW==0);iz=iz//SD;iy=iy//SH;ix=ix//SW
else:
iz=oz[:,None]*SD-PD+z[None,:]*DD;iy=oy[:,None]*SH-PH+y[None,:]*DH;ix=ox[:,None]*SW-PW+x[None,:]*DW;aligned=tl.full((BM,BK),True,tl.int1)
xo=batch[:,None]*X0+(group*CG+ic[None,:])*X1+iz*X2+iy*X3+ix*X4
a=tl.load(X+xo,(batch[:,None]<B)&(ic[None,:]<CG)&(iz>=0)&(iz<ID)&(iy>=0)&(iy<IH)&(ix>=0)&(ix<IW)&aligned,0)
if TRANSPOSED:wo=(group*CG+ic[:,None])*W0+n[None,:]*W1+z[:,None]*W2+y[:,None]*W3+x[:,None]*W4
else:wo=(group*OG+n[None,:])*W0+ic[:,None]*W1+z[:,None]*W2+y[:,None]*W3+x[:,None]*W4
w=tl.load(W+wo,(ic[:,None]<CG)&(n[None,:]<OG),0)
acc+=tl.dot(a,w,input_precision='tf32x3')
if HAS_BIAS:acc+=tl.load(Bias+(group*OG+n)*B0,n<OG,0)[None,:]
out=batch[:,None]*CO*volume+(group*OG+n[None,:])*volume+spatial[:,None]
tl.store(Y+out,acc,(batch[:,None]<B)&(n[None,:]<OG))
@triton.jit
def _ic_depthwise(X,W,Bias,Y,B:tl.constexpr,CI:tl.constexpr,CO:tl.constexpr,OG:tl.constexpr,
ID:tl.constexpr,IH:tl.constexpr,IW:tl.constexpr,OD:tl.constexpr,OH:tl.constexpr,OW:tl.constexpr,
KD:tl.constexpr,KH:tl.constexpr,KW:tl.constexpr,SD:tl.constexpr,SH:tl.constexpr,SW:tl.constexpr,
PD:tl.constexpr,PH:tl.constexpr,PW:tl.constexpr,DD:tl.constexpr,DH:tl.constexpr,DW:tl.constexpr,
X0:tl.constexpr,X1:tl.constexpr,X2:tl.constexpr,X3:tl.constexpr,X4:tl.constexpr,
W0:tl.constexpr,W1:tl.constexpr,W2:tl.constexpr,W3:tl.constexpr,W4:tl.constexpr,
B0:tl.constexpr,HAS_BIAS:tl.constexpr,TRANSPOSED:tl.constexpr,BLOCK:tl.constexpr):
i=(tl.program_id(0)*BLOCK+tl.arange(0,BLOCK)).to(tl.int64);volume=OD*OH*OW;batch=i//(CO*volume);co=i//volume%CO;group=co//OG
spatial=i%volume;oz=spatial//(OH*OW);oy=spatial//OW%OH;ox=spatial%OW;acc=tl.full((BLOCK,),0,tl.float32)
for k in range(KD*KH*KW):
z=k//(KH*KW);y=k//KW%KH;x=k%KW
if TRANSPOSED:
iz=oz+PD-z*DD;iy=oy+PH-y*DH;ix=ox+PW-x*DW;aligned=(iz%SD==0)&(iy%SH==0)&(ix%SW==0);iz=iz//SD;iy=iy//SH;ix=ix//SW
else:iz=oz*SD-PD+z*DD;iy=oy*SH-PH+y*DH;ix=ox*SW-PW+x*DW;aligned=tl.full((BLOCK,),True,tl.int1)
a=tl.load(X+batch*X0+group*X1+iz*X2+iy*X3+ix*X4,(batch<B)&(iz>=0)&(iz<ID)&(iy>=0)&(iy<IH)&(ix>=0)&(ix<IW)&aligned,0)
if TRANSPOSED:wo=group*W0+(co%OG)*W1+z*W2+y*W3+x*W4
else:wo=co*W0+z*W2+y*W3+x*W4
w=tl.load(W+wo,batch<B,0);acc=acc+a*w
if HAS_BIAS:acc+=tl.load(Bias+co*B0,batch<B,0)
tl.store(Y+i,acc,i<B*CO*volume)
def _ic_forward(x,module):
rank=len(module.kernel_size)
if x.ndim!=rank+2 or x.shape[1]!=module.in_channels or x.dtype!=torch.float32 or module.weight.dtype!=torch.float32 or not x.is_cuda:raise ValueError('FP32 CUDA batched convolution required')
if x.device!=module.weight.device or (module.bias is not None and module.bias.device!=x.device):raise ValueError('convolution parameters/device mismatch')
if module.bias is not None and module.bias.dtype!=torch.float32:raise ValueError('FP32 bias required')
if torch.is_grad_enabled() and x.requires_grad:raise ValueError('this seed implements forward only;autograd integration required')
if torch.cuda.get_device_capability(x.device)[0]<8:raise ValueError('Ampere-or-newer CUDA required')
dims=(1,)*(3-rank)+tuple(x.shape[2:]);kernel=(1,)*(3-rank)+module.kernel_size;stride=(1,)*(3-rank)+module.stride;padding=(0,)*(3-rank)+module.padding;dilation=(1,)*(3-rank)+module.dilation;op=(0,)*(3-rank)+module.output_padding
if any(v<=0 for v in dims):raise ValueError('positive input spatial dimensions required')
if module.transposed:out=tuple((v-1)*s-2*p+d*(k-1)+o+1 for v,s,p,d,k,o in zip(dims,stride,padding,dilation,kernel,op))
else:out=tuple((v+2*p-d*(k-1)-1)//s+1 for v,s,p,d,k in zip(dims,stride,padding,dilation,kernel))
if any(v<=0 for v in out):raise ValueError('invalid convolution output shape')
result=torch.empty((x.shape[0],module.out_channels,*out[-rank:]),device=x.device,dtype=torch.float32)
if result.numel()==0:return result
xs=(x.stride(0),x.stride(1))+(0,)*(3-rank)+tuple(x.stride()[2:]);w=module.weight;ws=(w.stride(0),w.stride(1))+(0,)*(3-rank)+tuple(w.stride()[2:]);cg=module.in_channels//module.groups;og=module.out_channels//module.groups
common=(x,w,module.bias if module.bias is not None else w,result,x.shape[0],module.in_channels,module.out_channels)
metadata=(*dims,*out,*kernel,*stride,*padding,*dilation,*xs,*ws,module.bias.stride(0) if module.bias is not None else 0,module.bias is not None,module.transposed)
with torch.cuda.device(x.device):
if cg==1:_ic_depthwise[(triton.cdiv(result.numel(),256),)](*common,og,*metadata,BLOCK=256,num_warps=4)
else:_ic_conv[(triton.cdiv(x.shape[0]*math.prod(out),16),triton.cdiv(og,32),module.groups)](*common,cg,og,*metadata,BM=16,BN=32,BK=32,num_warps=4)
return result
"""