1 | # For correctness, see test-correctness.sh.
|
---|
2 | # For performance:
|
---|
3 | # pushd ~/cfax
|
---|
4 | # . ~/setcfa build-fast
|
---|
5 | # popd
|
---|
6 | # make perfprogs CFA=$cfa -j8 MODE=performance
|
---|
7 | # make results-latest.csv RUN_DURATION_SEC=5 RUN_NUM_REPS=5 RUN_DATA_SIZE_MODE=common5
|
---|
8 | # cp results-latest.csv results-baseline.csv
|
---|
9 | # make results-latest.csv OP_MOVEMENTS=stack OP_POLARITIES=insfirst OP_ACCESSORS=allhead RUN_DURATION_SEC=5 RUN_NUM_REPS=5 RUN_DATA_SIZE_MODE=thorough
|
---|
10 | # cp results-latest.csv results-sizing.csv
|
---|
11 |
|
---|
12 | CFA = cfa
|
---|
13 | CXX = g++-11
|
---|
14 | UXX = ~/u++/u++-7.0.0/bin/u++
|
---|
15 |
|
---|
16 | MODE=performance
|
---|
17 | EXTRA_COMP_FLAGS=
|
---|
18 | RUN_NUM_REPS=3
|
---|
19 | RUN_DATA_SIZE_MODE=none
|
---|
20 | RUN_DURATION_SEC=5
|
---|
21 | RUN_TASKSET_CPULIST=6
|
---|
22 |
|
---|
23 | ifeq "$(MODE)" "performance"
|
---|
24 | PERFFLAGS_CFA = -DNDEBUG -O3 -nodebug
|
---|
25 | PERFFLAGS_CC = -DNDEBUG -O3
|
---|
26 | else ifeq "$(MODE)" "correctness"
|
---|
27 | PERFFLAGS_CFA = -O0 -g -debug
|
---|
28 | PERFFLAGS_CC = -O0 -g
|
---|
29 | else
|
---|
30 | $(error Bad MODE ($(MODE)); should be performance or correctness)
|
---|
31 | endif
|
---|
32 |
|
---|
33 | PERFFLAGS_CXX = $(PERFFLAGS_CC)
|
---|
34 | PERFFLAGS_UXX = $(PERFFLAGS_CFA)
|
---|
35 |
|
---|
36 | CFLAGS=$(PERFFLAGS_CC) $(EXTRA_COMP_FLAGS)
|
---|
37 |
|
---|
38 | SHELL = /usr/bin/bash
|
---|
39 |
|
---|
40 | # function: project an element from a filename that contains a delimited tuple
|
---|
41 | # (call proj,-,a-b-c.hfa,3)
|
---|
42 | # is
|
---|
43 | # c
|
---|
44 | define proj
|
---|
45 | $(word $(3),$(subst $(1), ,$(basename $(2))))
|
---|
46 | endef
|
---|
47 |
|
---|
48 | # functions: cross two lists, adding given delimiter between
|
---|
49 | # (call cross,-,a b c,1 2)
|
---|
50 | # is
|
---|
51 | # a-1 a-2 b-1 b-2 c-1 c-2
|
---|
52 | define cross
|
---|
53 | $(foreach x,$(2),$(foreach xs,$(3),$(x)$(1)$(xs)))
|
---|
54 | endef
|
---|
55 | define cross3
|
---|
56 | $(call cross,$(1),$(2),$(call cross,$(1),$(3),$(4)))
|
---|
57 | endef
|
---|
58 | define cross4
|
---|
59 | $(call cross,$(1),$(2),$(call cross3,$(1),$(3),$(4),$(5)))
|
---|
60 | endef
|
---|
61 | define cross5
|
---|
62 | $(call cross,$(1),$(2),$(call cross4,$(1),$(3),$(4),$(5),$(6)))
|
---|
63 | endef
|
---|
64 |
|
---|
65 | OP_MOVEMENTS=stack queue
|
---|
66 | OP_POLARITIES=insfirst inslast
|
---|
67 | OP_ACCESSORS=allhead inselem remelem
|
---|
68 | FX_SOLUTIONS=lq-tailq lq-list cfa-cfa upp-upp cpp-stlref
|
---|
69 |
|
---|
70 | OPS=$(call cross3,-,$(OP_MOVEMENTS),$(OP_POLARITIES),$(OP_ACCESSORS))
|
---|
71 | FXS=$(FX_SOLUTIONS)
|
---|
72 |
|
---|
73 | all : perfprogs results-latest.csv
|
---|
74 |
|
---|
75 | # Want to add functional dependency:
|
---|
76 | # if current FX_SOLUTION is lq-list then
|
---|
77 | # current OP_MOVEMENT must be stack and
|
---|
78 | # current OP_POLARITY must be insfirst
|
---|
79 | LQ_LIST_INCOMPAT_OP_MOVEMENTS=$(filter-out stack,$(OP_MOVEMENTS))
|
---|
80 | LQ_LIST_INCOMPAT_OP_POLARITIES=$(filter-out insfirst,$(OP_POLARITIES))
|
---|
81 | LQ_LIST_INCOMPAT_OPS=$(call cross3,-,$(LQ_LIST_INCOMPAT_OP_MOVEMENTS),$(OP_POLARITIES),$(OP_ACCESSORS)) \
|
---|
82 | $(call cross3,-,$(OP_MOVEMENTS),$(LQ_LIST_INCOMPAT_OP_POLARITIES),$(OP_ACCESSORS))
|
---|
83 | INCOMPAT=$(call cross,--,lq-list,$(LQ_LIST_INCOMPAT_OPS))
|
---|
84 | define filterFds
|
---|
85 | $(filter-out $(INCOMPAT),$(1))
|
---|
86 | endef
|
---|
87 |
|
---|
88 |
|
---|
89 | CORES_FULL=$(call cross,--,$(FXS),$(OPS)) # lq-tailq--stack-inslast-allhead
|
---|
90 | CORES=$(call filterFds,$(CORES_FULL)) # lq-tailq--stack-inslast-allhead
|
---|
91 |
|
---|
92 |
|
---|
93 | PERFPROGS=$(call cross,--,perfexp,$(CORES))
|
---|
94 |
|
---|
95 | perfprogs : $(PERFPROGS)
|
---|
96 |
|
---|
97 | perfexp--% driver--%.o result--%.1csv : FX=$(call proj,--,$@,2)
|
---|
98 | perfexp--% driver--%.o result--%.1csv : FX_COARSE=$(call proj,-,$(FX),1)
|
---|
99 | perfexp--% driver--%.o result--%.1csv : OP=$(call proj,--,$@,3)
|
---|
100 | perfexp--% driver--%.o result--%.1csv : OP_MOVEMENT=$(call proj,-,$(OP),1)
|
---|
101 | perfexp--% driver--%.o result--%.1csv : OP_POLARITY=$(call proj,-,$(OP),2)
|
---|
102 | perfexp--% driver--%.o result--%.1csv : OP_ACCESSOR=$(call proj,-,$(OP),3)
|
---|
103 | perfexp--% driver--%.o result--%.1csv : OP_DEFINES=-DOP_MOVEMENT=$(OP_MOVEMENT) -DOP_POLARITY=$(OP_POLARITY) -DOP_ACCESSOR=$(OP_ACCESSOR)
|
---|
104 |
|
---|
105 | perfexp--cfa-% driver--cfa-%.o : COMPILER=$(CFA) $(PERFFLAGS_CFA)
|
---|
106 | perfexp--lq-% driver--lq-%.o : COMPILER=$(CC) $(PERFFLAGS_CC)
|
---|
107 | perfexp--cpp-% driver--cpp-%.o : COMPILER=$(CXX) $(PERFFLAGS_CXX)
|
---|
108 | perfexp--upp-% driver--upp-%.o : COMPILER=$(UXX) $(PERFFLAGS_UXX)
|
---|
109 | perfexp--% driver--%.o : COMPILER=NO-COMPILER-FOR-$(FX_COARSE)
|
---|
110 |
|
---|
111 | # Without this %.d rule, ordinary make runs have noise about the recipe for driver--%.o being ill-formed when called on a *.d.
|
---|
112 | # https://stackoverflow.com/questions/3714041/why-does-this-makefile-execute-a-target-on-make-clean
|
---|
113 | # Whatever you -include gets called as a target first.
|
---|
114 | # One such example is driver--upp-upp--stack-insfirst-allhead.d
|
---|
115 | # Without this %.d rule, `make make driver--upp-upp--stack-insfirst-allhead.d` leads to the rule for driver--%.o firing.
|
---|
116 | # Though my dumb human eyes don't see the pattern as matching.
|
---|
117 | %.d:
|
---|
118 | @touch $@
|
---|
119 |
|
---|
120 | perfexp--% : driver--%.o observation.o
|
---|
121 | $(COMPILER) $(EXTRA_COMP_FLAGS) $^ -o $@
|
---|
122 |
|
---|
123 | driver--%.o : driver.c
|
---|
124 | $(COMPILER) $(EXTRA_COMP_FLAGS) -c $< $(OP_DEFINES) -include op-$(OP).h -include fx-$(FX).h -o $@ -MMD
|
---|
125 |
|
---|
126 | sayhi:
|
---|
127 | echo $(PERFPROGS)
|
---|
128 |
|
---|
129 |
|
---|
130 | ifeq "$(RUN_DATA_SIZE_MODE)" "common5"
|
---|
131 | RUN_DATA_SIZES=\
|
---|
132 | 7-1000000 \
|
---|
133 | 71-100000 \
|
---|
134 | 809-10000 \
|
---|
135 | 9051-1000 \
|
---|
136 | 72421-100
|
---|
137 | else ifeq "$(RUN_DATA_SIZE_MODE)" "thorough"
|
---|
138 | RUN_DATA_SIZES=\
|
---|
139 | 7-1000000 \
|
---|
140 | 11-100000 \
|
---|
141 | 13-100000 \
|
---|
142 | 19-100000 \
|
---|
143 | 29-100000 \
|
---|
144 | 37-100000 \
|
---|
145 | 53-100000 \
|
---|
146 | 71-100000 \
|
---|
147 | 101-10000 \
|
---|
148 | 149-10000 \
|
---|
149 | 211-10000 \
|
---|
150 | 283-10000 \
|
---|
151 | 401-10000 \
|
---|
152 | 569-10000 \
|
---|
153 | 809-10000 \
|
---|
154 | 1151-1000 \
|
---|
155 | 1601-1000 \
|
---|
156 | 2267-1000 \
|
---|
157 | 3203-1000 \
|
---|
158 | 4547-1000 \
|
---|
159 | 6473-1000 \
|
---|
160 | 9051-1000 \
|
---|
161 | 12809-100 \
|
---|
162 | 18119-100 \
|
---|
163 | 25601-100 \
|
---|
164 | 36209-100 \
|
---|
165 | 51203-100 \
|
---|
166 | 72421-100 \
|
---|
167 | 102407-10 \
|
---|
168 | 144817-10 \
|
---|
169 | 204803-10 \
|
---|
170 | 289637-10 \
|
---|
171 | 409609-10 \
|
---|
172 | 579263-10 \
|
---|
173 | 819229-10 \
|
---|
174 | 1158613-1 \
|
---|
175 | 1638431-1 \
|
---|
176 | 2317057-1 \
|
---|
177 | 3276803-1 \
|
---|
178 | 4634111-1 \
|
---|
179 | 6553621-1 \
|
---|
180 | 9268211-1
|
---|
181 | else ifeq "$(RUN_DATA_SIZE_MODE)" "manual"
|
---|
182 | ifeq "$(RUN_DATA_SIZES)" ""
|
---|
183 | $(error RUN_DATA_SIZE_MODE is manual but RUN_DATA_SIZES not given)
|
---|
184 | endif
|
---|
185 | else ifeq "$(RUN_DATA_SIZE_MODE)" "none"
|
---|
186 | # Assume user manages RUN_ARGS; empty RUN_ARGS just means run with compiled-in defaults
|
---|
187 | RUN_DATA_SIZES=none
|
---|
188 | else
|
---|
189 | $(error Bad RUN_DATA_SIZE_MODE ($(RUN_DATA_SIZE_MODE)); should be common5, thorough or manual)
|
---|
190 | endif
|
---|
191 |
|
---|
192 | RUN_REP_IDS=$(shell echo {1..$(RUN_NUM_REPS)}) # 1 2 3
|
---|
193 | RUN_REP_EXTS=$(call cross3,,run,$(RUN_REP_IDS),.1csv) # run1.1csv run2.1cav run3.1csv
|
---|
194 |
|
---|
195 | RUN_LAUNCHES=$(call cross,--,$(RUN_DATA_SIZES),$(RUN_REP_EXTS))
|
---|
196 |
|
---|
197 |
|
---|
198 |
|
---|
199 | RESULT1S=$(call cross,.,$(CORES),$(RUN_LAUNCHES)) # lq-tailq--stack-inslast-allhead.run2.1csv
|
---|
200 |
|
---|
201 |
|
---|
202 | RESULT1S_SHUFD=$(shell shuf -e $(RESULT1S))
|
---|
203 |
|
---|
204 | %.1csv : CORE=$(basename $(basename $@))
|
---|
205 | %.1csv : LAUNCH=$(subst .,,$(suffix $(basename $@)))
|
---|
206 | %.1csv : SIZING=$(call proj,--,$(LAUNCH),1)
|
---|
207 | %.1csv : NUMNODES=$(call proj,-,$(SIZING),1)
|
---|
208 | %.1csv : CHECKDONE=$(call proj,-,$(SIZING),2)
|
---|
209 | %.1csv : REP_ID=$(subst run,,$(call proj,--,$(LAUNCH),2))
|
---|
210 | %.1csv : RUN_ARGS=$(if $(filter none,$(SIZING)),,$(RUN_DURATION_SEC) $(CHECKDONE) $(NUMNODES) -1 $(REP_ID)) # use REP_ID as seed
|
---|
211 | %.1csv : REP_TIME=$(shell date '+%F %H:%M:%S')
|
---|
212 | %.1csv : perfprogs FORCE
|
---|
213 | taskset --cpu-list $(RUN_TASKSET_CPULIST) ./perfexp--$(CORE) $(RUN_ARGS) | xargs -n 1 printf '%s,%s,%s,%s\n' "$(REP_TIME)" "$(REP_ID)" "$(RUN_ARGS)" | tee $@
|
---|
214 |
|
---|
215 |
|
---|
216 | BATCHTIME=$(shell date '+%F--%H-%M-%S')
|
---|
217 |
|
---|
218 | results--$(BATCHTIME).csv : $(RESULT1S_SHUFD)
|
---|
219 | cat $^ | tee $@
|
---|
220 | rm $^
|
---|
221 |
|
---|
222 | results-latest.csv : results--$(BATCHTIME).csv
|
---|
223 | rm -f $@
|
---|
224 | ln -s $< $@
|
---|
225 |
|
---|
226 |
|
---|
227 |
|
---|
228 | clean :
|
---|
229 | rm -f *.o *.d perfexp--*
|
---|
230 |
|
---|
231 | # The FORCE business means any target that mentions it is also phony
|
---|
232 | # https://www.gnu.org/software/make/manual/html_node/Phony-Targets.html says: To always rebuild a pattern rule consider using a “force target” (see [https://www.gnu.org/software/make/manual/html_node/Force-Targets.html]).
|
---|
233 | .PHONY: all perfprogs results-latest.csv clean
|
---|
234 | FORCE:
|
---|
235 |
|
---|
236 | .PRECIOUS: result--%.1csv driver--%.o perfexp--% %.o
|
---|
237 |
|
---|
238 | -include *.d
|
---|