A Modern GPGPU API & wip linux RDNA2+ Driver
rdna
driver
linux
gpu
api: added compute shader dispatch
+62
+19
docs/drivers/amd/index.rst
+19
docs/drivers/amd/index.rst
···
23
23
- *IB*: Indirect Buffer: Allocated on BO, can be submitted to a HW_IP.
24
24
- *CS*: Command Stream
25
25
26
+
27
+
AMD (& afaik NV) support predication in multiple ways. In AMD, we can both skip over commands in the CP, but also
28
+
many packets can read from the predication register. Skipping commands is done for arbitrary va atomic ops (i believe),
29
+
and the register is for stuff like DRAW_VISIBLE. I believe it would be awesome to support an api like::
30
+
31
+
auto x = kes_malloc(dev, size, 4, KesMemoryDefault);
32
+
auto pred = kes_malloc(dev, 1, 4, KesMemoryDefault);
33
+
34
+
auto l1 = kes_start_recording(compute);
35
+
36
+
kes_cmd_conditional_begin(l1, pred, KesCondOpEqual);
37
+
kes_cmd_memset(l1, x.gpu, size, 2);
38
+
kes_cmd_memcpy(l1, y.gpu, x.gpu, size);
39
+
kes_cmd_conditional_end(l1);
40
+
41
+
kes_submit(compute, l1);
42
+
43
+
This would be awesome, as it would actually expose this stuff without weird extensions like VK_conditional_rendering.
44
+
26
45
.. toctree::
27
46
bugs
28
47
+2
kestrel/include/kestrel/interface.h
+2
kestrel/include/kestrel/interface.h
···
29
29
void (*fn_cmd_write_timestamp)(KesCommandList, kes_gpuptr_t addr);
30
30
void (*fn_cmd_signal_after)(KesCommandList, enum KesStage before, kes_gpuptr_t addr, uint64_t value, enum KesSignal);
31
31
void (*fn_cmd_wait_before)(KesCommandList, enum KesStage after, kes_gpuptr_t addr, uint64_t value, enum KesOp, enum KesHazardFlags, uint64_t mask);
32
+
void (*fn_cmd_dispatch)(KesCommandList command_list, uint32_t x, uint32_t y, uint32_t z);
33
+
void (*fn_cmd_dispatch_indirect)(KesCommandList command_list, kes_gpuptr_t command_addr);
32
34
};
33
35
34
36
/**
+27
kestrel/include/kestrel/kestrel.h
+27
kestrel/include/kestrel/kestrel.h
···
280
280
*/
281
281
void kes_cmd_wait_before(KesCommandList command_list, enum KesStage after, kes_gpuptr_t addr, uint64_t value, enum KesOp op, enum KesHazardFlags hazard_flags, uint64_t mask);
282
282
283
+
/**
284
+
* Contains arguments for a compute shader dispatch.
285
+
*/
286
+
struct KesDispatchIndirectCommand {
287
+
uint32_t x;
288
+
uint32_t y;
289
+
uint32_t z;
290
+
};
291
+
292
+
/**
293
+
* Record a compute shader dispatch in the command list.
294
+
* @param command_list
295
+
* @param x
296
+
* @param y
297
+
* @param z
298
+
*/
299
+
void kes_cmd_dispatch(KesCommandList command_list, uint32_t x, uint32_t y, uint32_t z);
300
+
301
+
/**
302
+
* Record an indirect compute shader dispatch in the command list.
303
+
* @param command_list The command list to record the command in.
304
+
* @param command_addr A GPU pointer to a KesDispatchIndirectCommand
305
+
* @sa KesDispatchIndirectCommand
306
+
*/
307
+
void kes_cmd_dispatch_indirect(KesCommandList command_list, kes_gpuptr_t command_addr);
308
+
309
+
283
310
#ifdef __cplusplus
284
311
}
285
312
#endif
+14
kestrel/rt/api.cpp
+14
kestrel/rt/api.cpp
···
200
200
201
201
dev->fns.fn_cmd_wait_before(clhandle->cmdlist, after, addr, value, op, hazard, mask);
202
202
}
203
+
204
+
API_EXPORT void kes_cmd_dispatch(KesCommandList pcl, uint32_t x, uint32_t y, uint32_t z) {
205
+
auto *clhandle = reinterpret_cast<CommandListHandle *>(pcl);
206
+
auto *dev = clhandle->dev;
207
+
208
+
dev->fns.fn_cmd_dispatch(clhandle->cmdlist, x, y, z);
209
+
}
210
+
211
+
API_EXPORT void kes_cmd_dispatch_indirect(KesCommandList pcl, kes_gpuptr_t command_addr) {
212
+
auto *clhandle = reinterpret_cast<CommandListHandle *>(pcl);
213
+
auto *dev = clhandle->dev;
214
+
215
+
dev->fns.fn_cmd_dispatch_indirect(clhandle->cmdlist, command_addr);
216
+
}