xref: /openbmc/qemu/hw/misc/macio/cuda.c (revision 13b5ae94)
1 /*
2  * QEMU PowerMac CUDA device support
3  *
4  * Copyright (c) 2004-2007 Fabrice Bellard
5  * Copyright (c) 2007 Jocelyn Mayer
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a copy
8  * of this software and associated documentation files (the "Software"), to deal
9  * in the Software without restriction, including without limitation the rights
10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11  * copies of the Software, and to permit persons to whom the Software is
12  * furnished to do so, subject to the following conditions:
13  *
14  * The above copyright notice and this permission notice shall be included in
15  * all copies or substantial portions of the Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23  * THE SOFTWARE.
24  */
25 
26 #include "qemu/osdep.h"
27 #include "hw/irq.h"
28 #include "hw/qdev-properties.h"
29 #include "migration/vmstate.h"
30 #include "hw/misc/macio/cuda.h"
31 #include "qemu/timer.h"
32 #include "sysemu/runstate.h"
33 #include "sysemu/rtc.h"
34 #include "qapi/error.h"
35 #include "qemu/cutils.h"
36 #include "qemu/log.h"
37 #include "qemu/module.h"
38 #include "trace.h"
39 
40 /* Bits in B data register: all active low */
41 #define TREQ            0x08    /* Transfer request (input) */
42 #define TACK            0x10    /* Transfer acknowledge (output) */
43 #define TIP             0x20    /* Transfer in progress (output) */
44 
45 /* commands (1st byte) */
46 #define ADB_PACKET      0
47 #define CUDA_PACKET     1
48 #define ERROR_PACKET    2
49 #define TIMER_PACKET    3
50 #define POWER_PACKET    4
51 #define MACIIC_PACKET   5
52 #define PMU_PACKET      6
53 
54 #define CUDA_TIMER_FREQ (4700000 / 6)
55 
56 /* CUDA returns time_t's offset from Jan 1, 1904, not 1970 */
57 #define RTC_OFFSET                      2082844800
58 
59 static void cuda_receive_packet_from_host(CUDAState *s,
60                                           const uint8_t *data, int len);
61 
62 /* MacOS uses timer 1 for calibration on startup, so we use
63  * the timebase frequency and cuda_get_counter_value() with
64  * cuda_get_load_time() to steer MacOS to calculate calibrate its timers
65  * correctly for both TCG and KVM (see commit b981289c49 "PPC: Cuda: Use cuda
66  * timer to expose tbfreq to guest" for more information) */
67 
68 static uint64_t cuda_get_counter_value(MOS6522State *s, MOS6522Timer *ti)
69 {
70     MOS6522CUDAState *mcs = container_of(s, MOS6522CUDAState, parent_obj);
71     CUDAState *cs = container_of(mcs, CUDAState, mos6522_cuda);
72 
73     /* Reverse of the tb calculation algorithm that Mac OS X uses on bootup */
74     uint64_t tb_diff = muldiv64(qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL),
75                                 cs->tb_frequency, NANOSECONDS_PER_SECOND) -
76                            ti->load_time;
77 
78     return (tb_diff * 0xBF401675E5DULL) / (cs->tb_frequency << 24);
79 }
80 
81 static uint64_t cuda_get_load_time(MOS6522State *s, MOS6522Timer *ti)
82 {
83     MOS6522CUDAState *mcs = container_of(s, MOS6522CUDAState, parent_obj);
84     CUDAState *cs = container_of(mcs, CUDAState, mos6522_cuda);
85 
86     uint64_t load_time = muldiv64(qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL),
87                                   cs->tb_frequency, NANOSECONDS_PER_SECOND);
88     return load_time;
89 }
90 
91 static void cuda_set_sr_int(void *opaque)
92 {
93     CUDAState *s = opaque;
94     MOS6522CUDAState *mcs = &s->mos6522_cuda;
95     MOS6522State *ms = MOS6522(mcs);
96     qemu_irq irq = qdev_get_gpio_in(DEVICE(ms), SR_INT_BIT);
97 
98     qemu_set_irq(irq, 1);
99 }
100 
101 static void cuda_delay_set_sr_int(CUDAState *s)
102 {
103     int64_t expire;
104 
105     trace_cuda_delay_set_sr_int();
106 
107     expire = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + s->sr_delay_ns;
108     timer_mod(s->sr_delay_timer, expire);
109 }
110 
111 /* NOTE: TIP and TREQ are negated */
112 static void cuda_update(CUDAState *s)
113 {
114     MOS6522CUDAState *mcs = &s->mos6522_cuda;
115     MOS6522State *ms = MOS6522(mcs);
116     ADBBusState *adb_bus = &s->adb_bus;
117     int packet_received, len;
118 
119     packet_received = 0;
120     if (!(ms->b & TIP)) {
121         /* transfer requested from host */
122 
123         if (ms->acr & SR_OUT) {
124             /* data output */
125             if ((ms->b & (TACK | TIP)) != (s->last_b & (TACK | TIP))) {
126                 if (s->data_out_index < sizeof(s->data_out)) {
127                     if (s->data_out_index == 0) {
128                         adb_autopoll_block(adb_bus);
129                     }
130                     trace_cuda_data_send(ms->sr);
131                     s->data_out[s->data_out_index++] = ms->sr;
132                     cuda_delay_set_sr_int(s);
133                 }
134             }
135         } else {
136             if (s->data_in_index < s->data_in_size) {
137                 /* data input */
138                 if ((ms->b & (TACK | TIP)) != (s->last_b & (TACK | TIP))) {
139                     ms->sr = s->data_in[s->data_in_index++];
140                     trace_cuda_data_recv(ms->sr);
141                     /* indicate end of transfer */
142                     if (s->data_in_index >= s->data_in_size) {
143                         ms->b = (ms->b | TREQ);
144                         adb_autopoll_unblock(adb_bus);
145                     }
146                     cuda_delay_set_sr_int(s);
147                 }
148             }
149         }
150     } else {
151         /* no transfer requested: handle sync case */
152         if ((s->last_b & TIP) && (ms->b & TACK) != (s->last_b & TACK)) {
153             /* update TREQ state each time TACK change state */
154             if (ms->b & TACK) {
155                 ms->b = (ms->b | TREQ);
156             } else {
157                 ms->b = (ms->b & ~TREQ);
158             }
159             cuda_delay_set_sr_int(s);
160         } else {
161             if (!(s->last_b & TIP)) {
162                 /* handle end of host to cuda transfer */
163                 packet_received = (s->data_out_index > 0);
164                 /* always an IRQ at the end of transfer */
165                 cuda_delay_set_sr_int(s);
166             }
167             /* signal if there is data to read */
168             if (s->data_in_index < s->data_in_size) {
169                 ms->b = (ms->b & ~TREQ);
170             }
171         }
172     }
173 
174     s->last_acr = ms->acr;
175     s->last_b = ms->b;
176 
177     /* NOTE: cuda_receive_packet_from_host() can call cuda_update()
178        recursively */
179     if (packet_received) {
180         len = s->data_out_index;
181         s->data_out_index = 0;
182         cuda_receive_packet_from_host(s, s->data_out, len);
183     }
184 }
185 
186 static void cuda_send_packet_to_host(CUDAState *s,
187                                      const uint8_t *data, int len)
188 {
189     int i;
190 
191     trace_cuda_packet_send(len);
192     for (i = 0; i < len; i++) {
193         trace_cuda_packet_send_data(i, data[i]);
194     }
195 
196     memcpy(s->data_in, data, len);
197     s->data_in_size = len;
198     s->data_in_index = 0;
199     cuda_update(s);
200     cuda_delay_set_sr_int(s);
201 }
202 
203 static void cuda_adb_poll(void *opaque)
204 {
205     CUDAState *s = opaque;
206     ADBBusState *adb_bus = &s->adb_bus;
207     uint8_t obuf[ADB_MAX_OUT_LEN + 2];
208     int olen;
209 
210     olen = adb_poll(adb_bus, obuf + 2, adb_bus->autopoll_mask);
211     if (olen > 0) {
212         obuf[0] = ADB_PACKET;
213         obuf[1] = 0x40; /* polled data */
214         cuda_send_packet_to_host(s, obuf, olen + 2);
215     }
216 }
217 
218 /* description of commands */
219 typedef struct CudaCommand {
220     uint8_t command;
221     const char *name;
222     bool (*handler)(CUDAState *s,
223                     const uint8_t *in_args, int in_len,
224                     uint8_t *out_args, int *out_len);
225 } CudaCommand;
226 
227 static bool cuda_cmd_autopoll(CUDAState *s,
228                               const uint8_t *in_data, int in_len,
229                               uint8_t *out_data, int *out_len)
230 {
231     ADBBusState *adb_bus = &s->adb_bus;
232     bool autopoll;
233 
234     if (in_len != 1) {
235         return false;
236     }
237 
238     autopoll = (in_data[0] != 0) ? true : false;
239 
240     adb_set_autopoll_enabled(adb_bus, autopoll);
241     return true;
242 }
243 
244 static bool cuda_cmd_set_autorate(CUDAState *s,
245                                   const uint8_t *in_data, int in_len,
246                                   uint8_t *out_data, int *out_len)
247 {
248     ADBBusState *adb_bus = &s->adb_bus;
249 
250     if (in_len != 1) {
251         return false;
252     }
253 
254     /* we don't want a period of 0 ms */
255     /* FIXME: check what real hardware does */
256     if (in_data[0] == 0) {
257         return false;
258     }
259 
260     adb_set_autopoll_rate_ms(adb_bus, in_data[0]);
261     return true;
262 }
263 
264 static bool cuda_cmd_set_device_list(CUDAState *s,
265                                      const uint8_t *in_data, int in_len,
266                                      uint8_t *out_data, int *out_len)
267 {
268     ADBBusState *adb_bus = &s->adb_bus;
269     uint16_t mask;
270 
271     if (in_len != 2) {
272         return false;
273     }
274 
275     mask = (((uint16_t)in_data[0]) << 8) | in_data[1];
276 
277     adb_set_autopoll_mask(adb_bus, mask);
278     return true;
279 }
280 
281 static bool cuda_cmd_powerdown(CUDAState *s,
282                                const uint8_t *in_data, int in_len,
283                                uint8_t *out_data, int *out_len)
284 {
285     if (in_len != 0) {
286         return false;
287     }
288 
289     qemu_system_shutdown_request(SHUTDOWN_CAUSE_GUEST_SHUTDOWN);
290     return true;
291 }
292 
293 static bool cuda_cmd_reset_system(CUDAState *s,
294                                   const uint8_t *in_data, int in_len,
295                                   uint8_t *out_data, int *out_len)
296 {
297     if (in_len != 0) {
298         return false;
299     }
300 
301     qemu_system_reset_request(SHUTDOWN_CAUSE_GUEST_RESET);
302     return true;
303 }
304 
305 static bool cuda_cmd_set_file_server_flag(CUDAState *s,
306                                           const uint8_t *in_data, int in_len,
307                                           uint8_t *out_data, int *out_len)
308 {
309     if (in_len != 1) {
310         return false;
311     }
312 
313     qemu_log_mask(LOG_UNIMP,
314                   "CUDA: unimplemented command FILE_SERVER_FLAG %d\n",
315                   in_data[0]);
316     return true;
317 }
318 
319 static bool cuda_cmd_set_power_message(CUDAState *s,
320                                        const uint8_t *in_data, int in_len,
321                                        uint8_t *out_data, int *out_len)
322 {
323     if (in_len != 1) {
324         return false;
325     }
326 
327     qemu_log_mask(LOG_UNIMP,
328                   "CUDA: unimplemented command SET_POWER_MESSAGE %d\n",
329                   in_data[0]);
330     return true;
331 }
332 
333 static bool cuda_cmd_get_time(CUDAState *s,
334                               const uint8_t *in_data, int in_len,
335                               uint8_t *out_data, int *out_len)
336 {
337     uint32_t ti;
338 
339     if (in_len != 0) {
340         return false;
341     }
342 
343     ti = s->tick_offset + (qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL)
344                            / NANOSECONDS_PER_SECOND);
345     out_data[0] = ti >> 24;
346     out_data[1] = ti >> 16;
347     out_data[2] = ti >> 8;
348     out_data[3] = ti;
349     *out_len = 4;
350     return true;
351 }
352 
353 static bool cuda_cmd_set_time(CUDAState *s,
354                               const uint8_t *in_data, int in_len,
355                               uint8_t *out_data, int *out_len)
356 {
357     uint32_t ti;
358 
359     if (in_len != 4) {
360         return false;
361     }
362 
363     ti = (((uint32_t)in_data[0]) << 24) + (((uint32_t)in_data[1]) << 16)
364          + (((uint32_t)in_data[2]) << 8) + in_data[3];
365     s->tick_offset = ti - (qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL)
366                            / NANOSECONDS_PER_SECOND);
367     return true;
368 }
369 
370 static const CudaCommand handlers[] = {
371     { CUDA_AUTOPOLL, "AUTOPOLL", cuda_cmd_autopoll },
372     { CUDA_SET_AUTO_RATE, "SET_AUTO_RATE",  cuda_cmd_set_autorate },
373     { CUDA_SET_DEVICE_LIST, "SET_DEVICE_LIST", cuda_cmd_set_device_list },
374     { CUDA_POWERDOWN, "POWERDOWN", cuda_cmd_powerdown },
375     { CUDA_RESET_SYSTEM, "RESET_SYSTEM", cuda_cmd_reset_system },
376     { CUDA_FILE_SERVER_FLAG, "FILE_SERVER_FLAG",
377       cuda_cmd_set_file_server_flag },
378     { CUDA_SET_POWER_MESSAGES, "SET_POWER_MESSAGES",
379       cuda_cmd_set_power_message },
380     { CUDA_GET_TIME, "GET_TIME", cuda_cmd_get_time },
381     { CUDA_SET_TIME, "SET_TIME", cuda_cmd_set_time },
382 };
383 
384 static void cuda_receive_packet(CUDAState *s,
385                                 const uint8_t *data, int len)
386 {
387     uint8_t obuf[16] = { CUDA_PACKET, 0, data[0] };
388     int i, out_len = 0;
389 
390     for (i = 0; i < ARRAY_SIZE(handlers); i++) {
391         const CudaCommand *desc = &handlers[i];
392         if (desc->command == data[0]) {
393             trace_cuda_receive_packet_cmd(desc->name);
394             out_len = 0;
395             if (desc->handler(s, data + 1, len - 1, obuf + 3, &out_len)) {
396                 cuda_send_packet_to_host(s, obuf, 3 + out_len);
397             } else {
398                 qemu_log_mask(LOG_GUEST_ERROR,
399                               "CUDA: %s: wrong parameters %d\n",
400                               desc->name, len);
401                 obuf[0] = ERROR_PACKET;
402                 obuf[1] = 0x5; /* bad parameters */
403                 obuf[2] = CUDA_PACKET;
404                 obuf[3] = data[0];
405                 cuda_send_packet_to_host(s, obuf, 4);
406             }
407             return;
408         }
409     }
410 
411     qemu_log_mask(LOG_GUEST_ERROR, "CUDA: unknown command 0x%02x\n", data[0]);
412     obuf[0] = ERROR_PACKET;
413     obuf[1] = 0x2; /* unknown command */
414     obuf[2] = CUDA_PACKET;
415     obuf[3] = data[0];
416     cuda_send_packet_to_host(s, obuf, 4);
417 }
418 
419 static void cuda_receive_packet_from_host(CUDAState *s,
420                                           const uint8_t *data, int len)
421 {
422     int i;
423 
424     trace_cuda_packet_receive(len);
425     for (i = 0; i < len; i++) {
426         trace_cuda_packet_receive_data(i, data[i]);
427     }
428 
429     switch(data[0]) {
430     case ADB_PACKET:
431         {
432             uint8_t obuf[ADB_MAX_OUT_LEN + 3];
433             int olen;
434             olen = adb_request(&s->adb_bus, obuf + 2, data + 1, len - 1);
435             if (olen > 0) {
436                 obuf[0] = ADB_PACKET;
437                 obuf[1] = 0x00;
438                 cuda_send_packet_to_host(s, obuf, olen + 2);
439             } else {
440                 /* error */
441                 obuf[0] = ADB_PACKET;
442                 obuf[1] = -olen;
443                 obuf[2] = data[1];
444                 olen = 0;
445                 cuda_send_packet_to_host(s, obuf, olen + 3);
446             }
447         }
448         break;
449     case CUDA_PACKET:
450         cuda_receive_packet(s, data + 1, len - 1);
451         break;
452     }
453 }
454 
455 static uint64_t mos6522_cuda_read(void *opaque, hwaddr addr, unsigned size)
456 {
457     CUDAState *s = opaque;
458     MOS6522CUDAState *mcs = &s->mos6522_cuda;
459     MOS6522State *ms = MOS6522(mcs);
460 
461     addr = (addr >> 9) & 0xf;
462     return mos6522_read(ms, addr, size);
463 }
464 
465 static void mos6522_cuda_write(void *opaque, hwaddr addr, uint64_t val,
466                                unsigned size)
467 {
468     CUDAState *s = opaque;
469     MOS6522CUDAState *mcs = &s->mos6522_cuda;
470     MOS6522State *ms = MOS6522(mcs);
471 
472     addr = (addr >> 9) & 0xf;
473     mos6522_write(ms, addr, val, size);
474 }
475 
476 static const MemoryRegionOps mos6522_cuda_ops = {
477     .read = mos6522_cuda_read,
478     .write = mos6522_cuda_write,
479     .endianness = DEVICE_BIG_ENDIAN,
480     .valid = {
481         .min_access_size = 1,
482         .max_access_size = 1,
483     },
484 };
485 
486 static const VMStateDescription vmstate_cuda = {
487     .name = "cuda",
488     .version_id = 6,
489     .minimum_version_id = 6,
490     .fields = (const VMStateField[]) {
491         VMSTATE_STRUCT(mos6522_cuda.parent_obj, CUDAState, 0, vmstate_mos6522,
492                        MOS6522State),
493         VMSTATE_UINT8(last_b, CUDAState),
494         VMSTATE_UINT8(last_acr, CUDAState),
495         VMSTATE_INT32(data_in_size, CUDAState),
496         VMSTATE_INT32(data_in_index, CUDAState),
497         VMSTATE_INT32(data_out_index, CUDAState),
498         VMSTATE_BUFFER(data_in, CUDAState),
499         VMSTATE_BUFFER(data_out, CUDAState),
500         VMSTATE_UINT32(tick_offset, CUDAState),
501         VMSTATE_TIMER_PTR(sr_delay_timer, CUDAState),
502         VMSTATE_END_OF_LIST()
503     }
504 };
505 
506 static void cuda_reset(DeviceState *dev)
507 {
508     CUDAState *s = CUDA(dev);
509     ADBBusState *adb_bus = &s->adb_bus;
510 
511     s->data_in_size = 0;
512     s->data_in_index = 0;
513     s->data_out_index = 0;
514 
515     adb_set_autopoll_enabled(adb_bus, false);
516 }
517 
518 static void cuda_realize(DeviceState *dev, Error **errp)
519 {
520     CUDAState *s = CUDA(dev);
521     SysBusDevice *sbd;
522     ADBBusState *adb_bus = &s->adb_bus;
523     struct tm tm;
524 
525     if (!sysbus_realize(SYS_BUS_DEVICE(&s->mos6522_cuda), errp)) {
526         return;
527     }
528 
529     /* Pass IRQ from 6522 */
530     sbd = SYS_BUS_DEVICE(s);
531     sysbus_pass_irq(sbd, SYS_BUS_DEVICE(&s->mos6522_cuda));
532 
533     qemu_get_timedate(&tm, 0);
534     s->tick_offset = (uint32_t)mktimegm(&tm) + RTC_OFFSET;
535 
536     s->sr_delay_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, cuda_set_sr_int, s);
537     s->sr_delay_ns = 20 * SCALE_US;
538 
539     adb_register_autopoll_callback(adb_bus, cuda_adb_poll, s);
540 }
541 
542 static void cuda_init(Object *obj)
543 {
544     CUDAState *s = CUDA(obj);
545     SysBusDevice *sbd = SYS_BUS_DEVICE(obj);
546 
547     object_initialize_child(obj, "mos6522-cuda", &s->mos6522_cuda,
548                             TYPE_MOS6522_CUDA);
549 
550     memory_region_init_io(&s->mem, obj, &mos6522_cuda_ops, s, "cuda", 0x2000);
551     sysbus_init_mmio(sbd, &s->mem);
552 
553     qbus_init(&s->adb_bus, sizeof(s->adb_bus), TYPE_ADB_BUS,
554               DEVICE(obj), "adb.0");
555 }
556 
557 static Property cuda_properties[] = {
558     DEFINE_PROP_UINT64("timebase-frequency", CUDAState, tb_frequency, 0),
559     DEFINE_PROP_END_OF_LIST()
560 };
561 
562 static void cuda_class_init(ObjectClass *oc, void *data)
563 {
564     DeviceClass *dc = DEVICE_CLASS(oc);
565 
566     dc->realize = cuda_realize;
567     dc->reset = cuda_reset;
568     dc->vmsd = &vmstate_cuda;
569     device_class_set_props(dc, cuda_properties);
570     set_bit(DEVICE_CATEGORY_BRIDGE, dc->categories);
571 }
572 
573 static const TypeInfo cuda_type_info = {
574     .name = TYPE_CUDA,
575     .parent = TYPE_SYS_BUS_DEVICE,
576     .instance_size = sizeof(CUDAState),
577     .instance_init = cuda_init,
578     .class_init = cuda_class_init,
579 };
580 
581 static void mos6522_cuda_portB_write(MOS6522State *s)
582 {
583     MOS6522CUDAState *mcs = container_of(s, MOS6522CUDAState, parent_obj);
584     CUDAState *cs = container_of(mcs, CUDAState, mos6522_cuda);
585 
586     cuda_update(cs);
587 }
588 
589 static void mos6522_cuda_reset_hold(Object *obj, ResetType type)
590 {
591     MOS6522State *ms = MOS6522(obj);
592     MOS6522DeviceClass *mdc = MOS6522_GET_CLASS(ms);
593 
594     if (mdc->parent_phases.hold) {
595         mdc->parent_phases.hold(obj, type);
596     }
597 
598     ms->timers[0].frequency = CUDA_TIMER_FREQ;
599     ms->timers[1].frequency = (SCALE_US * 6000) / 4700;
600 }
601 
602 static void mos6522_cuda_class_init(ObjectClass *oc, void *data)
603 {
604     ResettableClass *rc = RESETTABLE_CLASS(oc);
605     MOS6522DeviceClass *mdc = MOS6522_CLASS(oc);
606 
607     resettable_class_set_parent_phases(rc, NULL, mos6522_cuda_reset_hold,
608                                        NULL, &mdc->parent_phases);
609     mdc->portB_write = mos6522_cuda_portB_write;
610     mdc->get_timer1_counter_value = cuda_get_counter_value;
611     mdc->get_timer2_counter_value = cuda_get_counter_value;
612     mdc->get_timer1_load_time = cuda_get_load_time;
613     mdc->get_timer2_load_time = cuda_get_load_time;
614 }
615 
616 static const TypeInfo mos6522_cuda_type_info = {
617     .name = TYPE_MOS6522_CUDA,
618     .parent = TYPE_MOS6522,
619     .instance_size = sizeof(MOS6522CUDAState),
620     .class_init = mos6522_cuda_class_init,
621 };
622 
623 static void cuda_register_types(void)
624 {
625     type_register_static(&mos6522_cuda_type_info);
626     type_register_static(&cuda_type_info);
627 }
628 
629 type_init(cuda_register_types)
630