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 | @register_impls
class LinalgFunctions(InterpreterFunctions):
@impl(linalg.GenericOp)
def run_generic(
self, interpreter: Interpreter, op: linalg.GenericOp, args: tuple[Any, ...]
) -> PythonValues:
if op.library_call is not None:
raise NotImplementedError(
"library_call not yet supported in linalg.generic interpreter"
)
inputs_count = len(op.inputs)
input_args = args[:inputs_count]
output_args = args[inputs_count:]
for arg in output_args:
assert isinstance(arg, ShapedArray)
output_args = cast(tuple[ShapedArray[float], ...], output_args)
if op.results:
# If there are results, they must be tensors, initialised with the
# `output_args`. If not, the results are stored in output_args directly.
outputs = tuple(arg.copy() for arg in output_args)
else:
outputs = output_args
loop_shaped_args = input_args + outputs
indexing_maps = op.get_indexing_maps()
output_indexing_maps = indexing_maps[inputs_count:]
loop_ranges = op.get_static_loop_ranges()
for indices in product(*(range(loop_range) for loop_range in loop_ranges)):
loop_args = tuple(
(
(cast(ShapedArray[Any], i)).load(indexing_map.eval(indices, ()))
if isinstance(i, ShapedArray)
else i
)
for i, indexing_map in zip(loop_shaped_args, indexing_maps, strict=True)
)
loop_results = interpreter.run_ssacfg_region(op.body, loop_args, "for_loop")
for res, indexing_map in zip(
loop_results, output_indexing_maps, strict=True
):
result_indices = indexing_map.eval(indices, ())
outputs[0].store(result_indices, res)
return outputs if op.results else ()
@impl_terminator(linalg.YieldOp)
def run_yield(
self, interpreter: Interpreter, op: linalg.YieldOp, args: tuple[Any, ...]
):
return ReturnedValues(args), ()
@impl(linalg.AddOp)
def run_add(
self, interpreter: Interpreter, op: linalg.AddOp, args: tuple[Any, ...]
) -> tuple[Any, ...]:
(lhs, rhs, res) = (args[0], args[1], args[2])
assert isinstance(lhs, ShapedArray)
assert isinstance(rhs, ShapedArray)
assert isinstance(res, ShapedArray)
lhs = cast(ShapedArray[float], lhs)
rhs = cast(ShapedArray[float], rhs)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
assert lhs.shape == rhs.shape == res.shape
for i in range(len(lhs.data)):
res.data_ptr[i] = lhs.data_ptr[i] + rhs.data_ptr[i]
if len(op.results) > 0:
return (res,)
return ()
@impl(linalg.FillOp)
def run_fill(
self, interpreter: Interpreter, op: linalg.FillOp, args: tuple[Any, ...]
) -> tuple[Any, ...]:
operand, res = args[0], args[1]
assert isinstance(operand, ShapedArray)
assert isinstance(res, ShapedArray)
operand = cast(ShapedArray[float], operand)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
for i in range(len(res.data)):
res.data_ptr[i] = operand.data_ptr[0]
if len(op.results) > 0:
return (res,)
return ()
@impl(linalg.MulOp)
def run_mul(
self, interpreter: Interpreter, op: linalg.MulOp, args: tuple[Any, ...]
) -> tuple[Any, ...]:
lhs, rhs, res = args[0], args[1], args[2]
assert isinstance(lhs, ShapedArray)
assert isinstance(rhs, ShapedArray)
assert isinstance(res, ShapedArray)
lhs = cast(ShapedArray[float], lhs)
rhs = cast(ShapedArray[float], rhs)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
assert lhs.shape == rhs.shape == res.shape
for i in range(len(lhs.data)):
res.data_ptr[i] = lhs.data_ptr[i] * rhs.data_ptr[i]
if len(op.results) > 0:
return (res,)
return ()
@impl(linalg.TransposeOp)
def run_transpose(
self, interpreter: Interpreter, op: linalg.TransposeOp, args: tuple[Any, ...]
) -> tuple[Any, ...]:
operand, res = args[0], args[1]
assert isinstance(operand, ShapedArray)
assert isinstance(res, ShapedArray)
operand = cast(ShapedArray[float], operand)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
assert len(operand.shape) == 2
assert len(res.shape) == 2
rows, cols = operand.shape
for i in range(rows):
for j in range(cols):
res.data_ptr[j * rows + i] = operand.data_ptr[i * cols + j]
if len(op.results) > 0:
return (res,)
return ()
@impl(linalg.MatmulOp)
def run_mat_mul(
self, interpreter: Interpreter, op: linalg.MatmulOp, args: tuple[Any, ...]
) -> tuple[Any, ...]:
lhs, rhs, res = args[0], args[1], args[2]
assert isinstance(lhs, ShapedArray)
assert isinstance(rhs, ShapedArray)
assert isinstance(res, ShapedArray)
lhs = cast(ShapedArray[float], lhs)
rhs = cast(ShapedArray[float], rhs)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
rows = lhs.shape[0]
cols = rhs.shape[1]
assert rows == cols
for i in range(rows):
for j in range(cols):
res.data_ptr[i * cols + j] = sum(
lhs.data_ptr[i * lhs.shape[1] + k]
* rhs.data_ptr[k * rhs.shape[1] + j]
for k in range(lhs.shape[1])
)
if len(op.results) > 0:
return (res,)
return ()
@impl(linalg.PoolingNchwMaxOp)
def run_pooling_nchw_max(
self,
interpreter: Interpreter,
op: linalg.PoolingNchwMaxOp,
args: tuple[Any, ...],
) -> tuple[Any, ...]:
input, kernel_filter, res = args[0], args[1], args[2]
assert isinstance(input, ShapedArray)
assert isinstance(kernel_filter, ShapedArray)
assert isinstance(res, ShapedArray)
input = cast(ShapedArray[float], input)
kernel_filter = cast(ShapedArray[float], kernel_filter)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
strides_type = op.strides.type
assert isinstance(strides_type, TensorType)
(strides_shape,) = strides_type.get_shape()
strides = op.strides.get_values()
if strides_shape != 2:
raise NotImplementedError("Only 2d max pooling supported")
m_height, m_width = input.shape[2:]
ky, kx = kernel_filter.shape[0], kernel_filter.shape[1]
# convert input into a numpy like array
input_data = [
[input.data[r * m_width + c] for c in range(m_width)]
for r in range(m_height)
]
output: list[float] = []
for k in range(0, m_height - ky + 1, strides[0]):
for l in range(0, m_width - kx + 1, strides[0]):
block_max_value = float("-inf")
for i in range(k, k + ky):
for j in range(l, l + kx):
block_max_value = max(block_max_value, input_data[i][j])
output.append(block_max_value)
for i in range(len(output)):
res.data_ptr[i] = output[i]
if len(op.results) > 0:
return (res,)
return ()
@impl(linalg.Conv2DNchwFchwOp)
def run_conv_2d_nchw_fchw(
self,
interpreter: Interpreter,
op: linalg.Conv2DNchwFchwOp,
args: tuple[Any, ...],
) -> tuple[Any, ...]:
input, kernel_filter, res = args[0], args[1], args[2]
assert isinstance(input, ShapedArray)
assert isinstance(kernel_filter, ShapedArray)
assert isinstance(res, ShapedArray)
input = cast(ShapedArray[float], input)
kernel_filter = cast(ShapedArray[float], kernel_filter)
res = cast(ShapedArray[float], res)
if not all(res.data_ptr[i] == 0.0 for i in range(len(res.data))):
raise NotImplementedError()
m_height, m_width = input.shape[2:]
ky, kx = kernel_filter.shape[2], kernel_filter.shape[3]
strides = op.strides.get_values()
# convert input into a numpy like array
input_data = [
[input.data[r * m_width + c] for c in range(m_width)]
for r in range(m_height)
]
# convert kernel into a numpy like array
kernel_data = [
[
kernel_filter.data[r * kernel_filter.shape[3] + c]
for c in range(kernel_filter.shape[3])
]
for r in range(kernel_filter.shape[2])
]
output: list[float] = []
for k in range(0, m_height - ky + 1, strides[0]):
for l in range(0, m_width - kx + 1, strides[0]):
conv_value: float = 0.0
for i in range(k, k + ky):
for j in range(l, l + kx):
conv_value += input_data[i][j] * kernel_data[i - k][j - l]
output.append(conv_value)
for i in range(len(output)):
res.data_ptr[i] = output[i]
if len(op.results) > 0:
return (res,)
return ()
|