Skip to content

Commit c8a03a1

Browse files
zxybazhLokiiiiii
zxybazh
authored and
Lokiiiiii
committed
Add cuda tags and unit test (apache#7410)
* Add cuda tags and unit test * Add missing space * Remove extra indent * Modify macro def position * Fix clang format * Fix clang format for set_config
1 parent ae70b88 commit c8a03a1

File tree

3 files changed

+283
-6
lines changed

3 files changed

+283
-6
lines changed

‎src/target/tag.cc‎

Lines changed: 260 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -21,6 +21,8 @@
2121
* \file src/target/target_tag.cc
2222
* \brief Target tag registry
2323
*/
24+
25+
#include <tvm/ir/expr.h>
2426
#include <tvm/runtime/registry.h>
2527
#include <tvm/target/tag.h>
2628
#include <tvm/target/target.h>
@@ -66,12 +68,264 @@ Target TargetTag::AddTag(String name, Map<String, ObjectRef> config, bool overri
6668
return Target(config);
6769
}
6870

69-
/********** Register Target tags **********/
71+
/********** Register Target tags ***********/
72+
73+
#define TVM_REGISTER_CUDA_TAG(Name, Arch, SharedMem, RegPerBlock) \
74+
TVM_REGISTER_TARGET_TAG(Name).set_config({ \
75+
{"kind", String("cuda")}, \
76+
{"arch", String(Arch)}, \
77+
{"shared_memory_per_block", Integer(SharedMem)}, \
78+
{"registers_per_block", Integer(RegPerBlock)}, \
79+
{"max_threads_per_block", Integer(1024)}, \
80+
{"thread_warp_size", Integer(32)}, \
81+
});
7082

71-
TVM_REGISTER_TARGET_TAG("nvidia/rtx2080ti")
72-
.set_config({
73-
{"kind", String("cuda")},
74-
{"arch", String("sm_75")},
75-
});
83+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k80", "sm_37", 49152, 65536);
84+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k40", "sm_35", 49152, 65536);
85+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k20", "sm_35", 49152, 65536);
86+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-c2075", "sm_20", 49152, 32768);
87+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-c2050", "sm_20", 49152, 32768);
88+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-c2070", "sm_20", 49152, 32768);
89+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-a100", "sm_80", 49152, 65536);
90+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-t4", "sm_75", 49152, 65536);
91+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-v100", "sm_70", 49152, 65536);
92+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-p100", "sm_60", 49152, 65536);
93+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-p40", "sm_61", 49152, 65536);
94+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-p4", "sm_61", 49152, 65536);
95+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-m60", "sm_52", 49152, 65536);
96+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-m40", "sm_52", 49152, 65536);
97+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k80", "sm_37", 49152, 65536);
98+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k40", "sm_35", 49152, 65536);
99+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k20", "sm_35", 49152, 65536);
100+
TVM_REGISTER_CUDA_TAG("nvidia/tesla-k10", "sm_30", 49152, 65536);
101+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-rtx-8000", "sm_75", 49152, 65536);
102+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-rtx-6000", "sm_75", 49152, 65536);
103+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-rtx-5000", "sm_75", 49152, 65536);
104+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-rtx-4000", "sm_75", 49152, 65536);
105+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-gv100", "sm_70", 49152, 65536);
106+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-gp100", "sm_60", 49152, 65536);
107+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p6000", "sm_61", 49152, 65536);
108+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p5000", "sm_61", 49152, 65536);
109+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p4000", "sm_61", 49152, 65536);
110+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p2200", "sm_61", 49152, 65536);
111+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p2000", "sm_61", 49152, 65536);
112+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p1000", "sm_61", 49152, 65536);
113+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p620", "sm_61", 49152, 65536);
114+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p600", "sm_61", 49152, 65536);
115+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p400", "sm_61", 49152, 65536);
116+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m6000-24gb", "sm_52", 49152, 65536);
117+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m6000", "sm_52", 49152, 65536);
118+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k6000", "sm_35", 49152, 65536);
119+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m5000", "sm_52", 49152, 65536);
120+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k5200", "sm_35", 49152, 65536);
121+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k5000", "sm_30", 49152, 65536);
122+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m4000", "sm_52", 49152, 65536);
123+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k4200", "sm_30", 49152, 65536);
124+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k4000", "sm_30", 49152, 65536);
125+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m2000", "sm_52", 49152, 65536);
126+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k2200", "sm_50", 49152, 65536);
127+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k2000", "sm_30", 49152, 65536);
128+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k2000d", "sm_30", 49152, 65536);
129+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k1200", "sm_50", 49152, 65536);
130+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k620", "sm_50", 49152, 65536);
131+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k600", "sm_30", 49152, 65536);
132+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k420", "sm_30", 49152, 65536);
133+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-410", "sm_30", 49152, 65536);
134+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-plex-7000", "sm_20", 49152, 32768);
135+
TVM_REGISTER_CUDA_TAG("nvidia/rtx-5000", "sm_75", 49152, 65536);
136+
TVM_REGISTER_CUDA_TAG("nvidia/rtx-4000", "sm_75", 49152, 65536);
137+
TVM_REGISTER_CUDA_TAG("nvidia/rtx-3000", "sm_75", 49152, 65536);
138+
TVM_REGISTER_CUDA_TAG("nvidia/t2000", "sm_75", 49152, 65536);
139+
TVM_REGISTER_CUDA_TAG("nvidia/t1000", "sm_75", 49152, 65536);
140+
TVM_REGISTER_CUDA_TAG("nvidia/p620", "sm_61", 49152, 65536);
141+
TVM_REGISTER_CUDA_TAG("nvidia/p520", "sm_61", 49152, 65536);
142+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p5200", "sm_61", 49152, 65536);
143+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p4200", "sm_61", 49152, 65536);
144+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p3200", "sm_61", 49152, 65536);
145+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p5000", "sm_61", 49152, 65536);
146+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p4000", "sm_61", 49152, 65536);
147+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p3000", "sm_61", 49152, 65536);
148+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p2000", "sm_61", 49152, 65536);
149+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p1000", "sm_61", 49152, 65536);
150+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p600", "sm_61", 49152, 65536);
151+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-p500", "sm_61", 49152, 65536);
152+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m5500m", "sm_52", 49152, 65536);
153+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m2200", "sm_52", 49152, 65536);
154+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m1200", "sm_50", 49152, 65536);
155+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m620", "sm_52", 49152, 65536);
156+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m520", "sm_50", 49152, 65536);
157+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k6000m", "sm_30", 49152, 65536);
158+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k5200m", "sm_30", 49152, 65536);
159+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k5100m", "sm_30", 49152, 65536);
160+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m5000m", "sm_50", 49152, 65536);
161+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k500m", "sm_30", 49152, 65536);
162+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k4200m", "sm_30", 49152, 65536);
163+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k4100m", "sm_30", 49152, 65536);
164+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m4000m", "sm_50", 49152, 65536);
165+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k3100m", "sm_30", 49152, 65536);
166+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m3000m", "sm_50", 49152, 65536);
167+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k2200m", "sm_30", 49152, 65536);
168+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k2100m", "sm_30", 49152, 65536);
169+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m2000m", "sm_50", 49152, 65536);
170+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k1100m", "sm_30", 49152, 65536);
171+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m1000m", "sm_50", 49152, 65536);
172+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k620m", "sm_50", 49152, 65536);
173+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k610m", "sm_35", 49152, 65536);
174+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m600m", "sm_50", 49152, 65536);
175+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-k510m", "sm_35", 49152, 65536);
176+
TVM_REGISTER_CUDA_TAG("nvidia/quadro-m500m", "sm_50", 49152, 65536);
177+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-nvs-810", "sm_50", 49152, 65536);
178+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-nvs-510", "sm_30", 49152, 65536);
179+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-nvs-315", "sm_21", 49152, 32768);
180+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-nvs-310", "sm_21", 49152, 32768);
181+
TVM_REGISTER_CUDA_TAG("nvidia/nvs-5400m", "sm_21", 49152, 32768);
182+
TVM_REGISTER_CUDA_TAG("nvidia/nvs-5200m", "sm_21", 49152, 32768);
183+
TVM_REGISTER_CUDA_TAG("nvidia/nvs-4200m", "sm_21", 49152, 32768);
184+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-3090", "sm_86", 49152, 65536);
185+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-3080", "sm_86", 49152, 65536);
186+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-3070", "sm_86", 49152, 65536);
187+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-titan-rtx", "sm_75", 49152, 65536);
188+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2080-ti", "sm_75", 49152, 65536);
189+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2080", "sm_75", 49152, 65536);
190+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2070", "sm_75", 49152, 65536);
191+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2060", "sm_75", 49152, 65536);
192+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-titan-v", "sm_70", 49152, 65536);
193+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-titan-xp", "sm_61", 49152, 65536);
194+
TVM_REGISTER_CUDA_TAG("nvidia/nvidia-titan-x", "sm_61", 49152, 65536);
195+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1080-ti", "sm_61", 49152, 65536);
196+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1080", "sm_61", 49152, 65536);
197+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1070-ti", "sm_61", 49152, 65536);
198+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1070", "sm_61", 49152, 65536);
199+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1060", "sm_61", 49152, 65536);
200+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1050", "sm_61", 49152, 65536);
201+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-titan-x", "sm_52", 49152, 65536);
202+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-titan-z", "sm_35", 49152, 65536);
203+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-titan-black", "sm_35", 49152, 65536);
204+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-titan", "sm_35", 49152, 65536);
205+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-980-ti", "sm_52", 49152, 65536);
206+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-980", "sm_52", 49152, 65536);
207+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-970", "sm_52", 49152, 65536);
208+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-960", "sm_52", 49152, 65536);
209+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-950", "sm_52", 49152, 65536);
210+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-780-ti", "sm_35", 49152, 65536);
211+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-780", "sm_35", 49152, 65536);
212+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-770", "sm_30", 49152, 65536);
213+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-760", "sm_30", 49152, 65536);
214+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-750-ti", "sm_50", 49152, 65536);
215+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-750", "sm_50", 49152, 65536);
216+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-690", "sm_30", 49152, 65536);
217+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-680", "sm_30", 49152, 65536);
218+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-670", "sm_30", 49152, 65536);
219+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-660-ti", "sm_30", 49152, 65536);
220+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-660", "sm_30", 49152, 65536);
221+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-650-ti-boost", "sm_30", 49152, 65536);
222+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-650-ti", "sm_30", 49152, 65536);
223+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-650", "sm_30", 49152, 65536);
224+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-560-ti", "sm_21", 49152, 32768);
225+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-550-ti", "sm_21", 49152, 32768);
226+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-460", "sm_21", 49152, 32768);
227+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gts-450", "sm_21", 49152, 32768);
228+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gts-450*", "sm_21", 49152, 32768);
229+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-590", "sm_20", 49152, 32768);
230+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-580", "sm_20", 49152, 32768);
231+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-570", "sm_20", 49152, 32768);
232+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-480", "sm_20", 49152, 32768);
233+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-470", "sm_20", 49152, 32768);
234+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-465", "sm_20", 49152, 32768);
235+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-740", "sm_30", 49152, 65536);
236+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-730", "sm_35", 49152, 65536);
237+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-730-ddr3,128bit", "sm_21", 49152, 32768);
238+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-720", "sm_35", 49152, 65536);
239+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-705*", "sm_35", 49152, 65536);
240+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-640-(gddr5)", "sm_35", 49152, 65536);
241+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-640-(gddr3)", "sm_21", 49152, 32768);
242+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-630", "sm_21", 49152, 32768);
243+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-620", "sm_21", 49152, 32768);
244+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-610", "sm_21", 49152, 32768);
245+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-520", "sm_21", 49152, 32768);
246+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-440", "sm_21", 49152, 32768);
247+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-440*", "sm_21", 49152, 32768);
248+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-430", "sm_21", 49152, 32768);
249+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-430*", "sm_21", 49152, 32768);
250+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2080", "sm_75", 49152, 65536);
251+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2070", "sm_75", 49152, 65536);
252+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-rtx-2060", "sm_75", 49152, 65536);
253+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1080", "sm_61", 49152, 65536);
254+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1070", "sm_61", 49152, 65536);
255+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-1060", "sm_61", 49152, 65536);
256+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-980", "sm_52", 49152, 65536);
257+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-980m", "sm_52", 49152, 65536);
258+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-970m", "sm_52", 49152, 65536);
259+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-965m", "sm_52", 49152, 65536);
260+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-960m", "sm_50", 49152, 65536);
261+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-950m", "sm_50", 49152, 65536);
262+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-940m", "sm_50", 49152, 65536);
263+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-930m", "sm_50", 49152, 65536);
264+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-920m", "sm_35", 49152, 65536);
265+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-910m", "sm_52", 49152, 65536);
266+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-880m", "sm_30", 49152, 65536);
267+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-870m", "sm_30", 49152, 65536);
268+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-860m-cuda-sm_30", "sm_30", 49152, 65536);
269+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-860m-sm-50", "sm_50", 49152, 65536);
270+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-850m", "sm_50", 49152, 65536);
271+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-840m", "sm_50", 49152, 65536);
272+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-830m", "sm_50", 49152, 65536);
273+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-820m", "sm_21", 49152, 32768);
274+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-800m", "sm_21", 49152, 32768);
275+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-780m", "sm_30", 49152, 65536);
276+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-770m", "sm_30", 49152, 65536);
277+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-765m", "sm_30", 49152, 65536);
278+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-760m", "sm_30", 49152, 65536);
279+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-680mx", "sm_30", 49152, 65536);
280+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-680m", "sm_30", 49152, 65536);
281+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-675mx", "sm_30", 49152, 65536);
282+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-675m", "sm_21", 49152, 32768);
283+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-670mx", "sm_30", 49152, 65536);
284+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-670m", "sm_21", 49152, 32768);
285+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-660m", "sm_30", 49152, 65536);
286+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-755m", "sm_30", 49152, 65536);
287+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-750m", "sm_30", 49152, 65536);
288+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-650m", "sm_30", 49152, 65536);
289+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-745m", "sm_30", 49152, 65536);
290+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-645m", "sm_30", 49152, 65536);
291+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-740m", "sm_30", 49152, 65536);
292+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-730m", "sm_30", 49152, 65536);
293+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-640m", "sm_30", 49152, 65536);
294+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-640m-le", "sm_30", 49152, 65536);
295+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-735m", "sm_30", 49152, 65536);
296+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-635m", "sm_21", 49152, 32768);
297+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-730m", "sm_30", 49152, 65536);
298+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-630m", "sm_21", 49152, 32768);
299+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-625m", "sm_21", 49152, 32768);
300+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-720m", "sm_21", 49152, 32768);
301+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-620m", "sm_21", 49152, 32768);
302+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-710m", "sm_21", 49152, 32768);
303+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-705m", "sm_21", 49152, 32768);
304+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-610m", "sm_21", 49152, 32768);
305+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-580m", "sm_21", 49152, 32768);
306+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-570m", "sm_21", 49152, 32768);
307+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-560m", "sm_21", 49152, 32768);
308+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-555m", "sm_21", 49152, 32768);
309+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-550m", "sm_21", 49152, 32768);
310+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-540m", "sm_21", 49152, 32768);
311+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-525m", "sm_21", 49152, 32768);
312+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-520mx", "sm_21", 49152, 32768);
313+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-520m", "sm_21", 49152, 32768);
314+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-485m", "sm_21", 49152, 32768);
315+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-470m", "sm_21", 49152, 32768);
316+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-460m", "sm_21", 49152, 32768);
317+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-445m", "sm_21", 49152, 32768);
318+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-435m", "sm_21", 49152, 32768);
319+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-420m", "sm_21", 49152, 32768);
320+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gt-415m", "sm_21", 49152, 32768);
321+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-gtx-480m", "sm_20", 49152, 32768);
322+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-710m", "sm_21", 49152, 32768);
323+
TVM_REGISTER_CUDA_TAG("nvidia/geforce-410m", "sm_21", 49152, 32768);
324+
TVM_REGISTER_CUDA_TAG("nvidia/jetson-agx-xavier", "sm_72", 49152, 65536);
325+
TVM_REGISTER_CUDA_TAG("nvidia/jetson-nano", "sm_53", 49152, 32768);
326+
TVM_REGISTER_CUDA_TAG("nvidia/jetson-tx2", "sm_62", 49152, 32768);
327+
TVM_REGISTER_CUDA_TAG("nvidia/jetson-tx1", "sm_53", 49152, 32768);
328+
TVM_REGISTER_CUDA_TAG("nvidia/tegra-x1", "sm_53", 49152, 32768);
76329

330+
#undef TVM_REGISTER_CUDA_TAG
77331
} // namespace tvm

‎src/target/target_kind.cc‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -230,6 +230,9 @@ TVM_REGISTER_TARGET_KIND("cuda", kDLGPU)
230230
.add_attr_option<Bool>("system-lib")
231231
.add_attr_option<Integer>("max_num_threads", Integer(1024))
232232
.add_attr_option<Integer>("thread_warp_size", Integer(32))
233+
.add_attr_option<Integer>("shared_memory_per_block")
234+
.add_attr_option<Integer>("registers_per_block")
235+
.add_attr_option<Integer>("max_threads_per_block")
233236
.set_default_keys({"cuda", "gpu"});
234237

235238
TVM_REGISTER_TARGET_KIND("nvptx", kDLGPU)

‎tests/python/unittest/test_target_target.py‎

Lines changed: 20 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -131,6 +131,26 @@ def test_composite_target():
131131
assert opencl_device.kind.name == "opencl"
132132

133133

134+
def test_target_tag_0():
135+
tgt = tvm.target.Target("nvidia/geforce-rtx-2080-ti")
136+
assert tgt.kind.name == "cuda"
137+
assert tgt.attrs["arch"] == "sm_75"
138+
assert tgt.attrs["shared_memory_per_block"] == 49152
139+
assert tgt.attrs["max_threads_per_block"] == 1024
140+
assert tgt.attrs["thread_warp_size"] == 32
141+
assert tgt.attrs["registers_per_block"] == 65536
142+
143+
144+
def test_target_tag_1():
145+
tgt = tvm.target.Target("nvidia/jetson-nano")
146+
assert tgt.kind.name == "cuda"
147+
assert tgt.attrs["arch"] == "sm_53"
148+
assert tgt.attrs["shared_memory_per_block"] == 49152
149+
assert tgt.attrs["max_threads_per_block"] == 1024
150+
assert tgt.attrs["thread_warp_size"] == 32
151+
assert tgt.attrs["registers_per_block"] == 32768
152+
153+
134154
if __name__ == "__main__":
135155
test_target_dispatch()
136156
test_target_string_parse()

0 commit comments

Comments
 (0)