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
|
include ../support/Makefile.inc
OPS = $(BIN)/$(HL_TARGET)/add_float32.a \
$(BIN)/$(HL_TARGET)/add_float64.a \
$(BIN)/$(HL_TARGET)/add_grad_float32.a \
$(BIN)/$(HL_TARGET)/add_grad_float64.a \
$(BIN)/$(HL_TARGET)/add_halidegrad_float32.a \
$(BIN)/$(HL_TARGET)/add_halidegrad_float64.a
# Check whether we have cuda installed, if add the CUDA ops as dependencies
ifeq ($(shell which nvcc),)
HAS_CUDA = 0
CUDA_OPS =
else
HAS_CUDA = 1
CUDA_OPS = $(BIN)/$(HL_TARGET)/add_cuda_float32.a \
$(BIN)/$(HL_TARGET)/add_cuda_float64.a \
$(BIN)/$(HL_TARGET)/add_grad_cuda_float32.a \
$(BIN)/$(HL_TARGET)/add_grad_cuda_float64.a \
$(BIN)/$(HL_TARGET)/add_halidegrad_cuda_float32.a \
$(BIN)/$(HL_TARGET)/add_halidegrad_cuda_float64.a
endif
EXT_LIB=$(BIN)/lib
PYTHON ?= python3
PIP ?= pip3
all: test
CUDA_TARGET ?= host-cuda-cuda_capability_61-user_context
# Run the PyTorch tests to verify the module is compiled correctly
# .wrapper is a dummy file whose timestamps allow Make to track the dependency
# on the Python side of the build
test: $(BIN)/.wrapper
@PYTHONPATH=$(EXT_LIB):${PYTHONPATH} $(PYTHON) test.py
# Build the python PyTorch extension that links against the Halide operators
$(BIN)/.wrapper: $(OPS) $(CUDA_OPS) setup.py
@mkdir -p $(EXT_LIB)
@HAS_CUDA=$(HAS_CUDA) \
HALIDE_DISTRIB_PATH=$(HALIDE_DISTRIB_PATH) \
BIN=$(BIN)/$(HL_TARGET) \
PYTHONPATH=$(EXT_LIB):${PYTHONPATH} \
$(PIP) install . -t $(EXT_LIB)
@touch $(BIN)/.wrapper
# Some common args
ADD_TYPES_F32 =\
input_a.type=float32 \
input_b.type=float32 \
output.type=float32 \
ADD_TYPES_F64 =\
input_a.type=float64 \
input_b.type=float64 \
output.type=float64 \
ADD_GRAD_TYPES_F32 =\
input_a.type=float32 \
input_b.type=float32 \
d_input_a.type=float32 \
d_input_b.type=float32 \
d_output.type=float32 \
ADD_GRAD_TYPES_F64 =\
input_a.type=float64 \
input_b.type=float64 \
d_input_a.type=float64 \
d_input_b.type=float64 \
d_output.type=float64 \
# Generate the CPU version of the op ------------------------------------------
$(BIN)/%/add_float32.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo Producing CPU operator
@$^ -g add \
$(ADD_TYPES_F32) \
-f add_float32 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$*
$(BIN)/%/add_halidegrad_float32.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo Producing CPU halidegrad
@$^ -g add \
$(ADD_TYPES_F32) \
-f add_halidegrad_float32 \
-e static_library,c_header,pytorch_wrapper \
-p $(HALIDE_DISTRIB_PATH)/lib/libautoschedule_li2018.so \
-o $(@D) \
-d 1 \
target=$* \
autoscheduler=Li2018
$(BIN)/%/add_grad_float32.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo Producing CPU gradient
@$^ -g add_grad \
$(ADD_GRAD_TYPES_F32) \
-f add_grad_float32 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$*
$(BIN)/%/add_float64.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo "Producing CPU (double) operator"
@$^ -g add \
$(ADD_TYPES_F64) \
-f add_float64 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$*
$(BIN)/%/add_halidegrad_float64.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo "Producing CPU (double) halidegrad"
@$^ -g add \
$(ADD_TYPES_F64) \
-f add_halidegrad_float64 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
-p $(HALIDE_DISTRIB_PATH)/lib/libautoschedule_li2018.so \
target=$* \
-d 1 \
target=$* \
autoscheduler=Li2018
$(BIN)/%/add_grad_float64.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo "Producing CPU (double) gradient"
@$^ -g add_grad \
$(ADD_GRAD_TYPES_F64) \
-f add_grad_float64 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$*
# -----------------------------------------------------------------------------
# Generate the GPU version of the op ------------------------------------------
$(BIN)/%/add_cuda_float32.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo Producing CUDA operator
@$^ -g add \
$(ADD_TYPES_F32) \
-f add_cuda_float32 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$(CUDA_TARGET)
$(BIN)/%/add_halidegrad_cuda_float32.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo Producing CUDA halidegrad
@$^ -g add \
$(ADD_TYPES_F32) \
-f add_halidegrad_cuda_float32 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
-p $(HALIDE_DISTRIB_PATH)/lib/libautoschedule_li2018.so \
-d 1 \
target=$(CUDA_TARGET) \
autoscheduler=Li2018
$(BIN)/%/add_grad_cuda_float32.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo "Producing CUDA (double) gradient"
@$^ -g add_grad \
$(ADD_GRAD_TYPES_F32) \
-f add_grad_cuda_float32 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$(CUDA_TARGET)
$(BIN)/%/add_cuda_float64.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo "Producing CUDA (double) operator"
@$^ -g add \
$(ADD_TYPES_F64) \
-f add_cuda_float64 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$(CUDA_TARGET)
$(BIN)/%/add_halidegrad_cuda_float64.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo "Producing CUDA (double) halidegrad"
@$^ -g add \
$(ADD_TYPES_F64) \
-f add_halidegrad_cuda_float64 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
-p $(HALIDE_DISTRIB_PATH)/lib/libautoschedule_li2018.so \
-d 1 \
target=$(CUDA_TARGET) \
autoscheduler=Li2018
$(BIN)/%/add_grad_cuda_float64.a: $(GENERATOR_BIN)/add.generator
@mkdir -p $(@D)
@echo Producing CUDA gradient
@$^ -g add_grad \
$(ADD_GRAD_TYPES_F64) \
-f add_grad_cuda_float64 \
-e static_library,c_header,pytorch_wrapper \
-o $(@D) \
target=$(CUDA_TARGET)
# -----------------------------------------------------------------------------
# Build the Halide generator for the operator
$(GENERATOR_BIN)/add.generator: src/add_generator.cpp $(GENERATOR_DEPS)
@echo Building Generator
@mkdir -p $(@D)
@$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) \
-o $@ $(LIBHALIDE_LDFLAGS)
clean:
rm -rf $(BIN) __pycache__
|
