A Modern GPGPU API & wip linux RDNA2+ Driver
rdna
driver
linux
gpu
api: added memcpy
+56
+2
kestrel/include/kestrel/interface.h
+2
kestrel/include/kestrel/interface.h
···
1
1
#pragma once
2
2
3
+
#include "kestrel.h"
3
4
#include "kestrel/kestrel.h"
4
5
5
6
#ifdef __cplusplus
···
24
25
KesCommandList (*fn_start_recording)(KesQueue);
25
26
void (*fn_submit)(KesQueue, KesCommandList);
26
27
void (*fn_cmd_memset)(KesCommandList, kes_gpuptr_t addr, size_t size, uint32_t value);
28
+
void (*fn_cmd_memcpy)(KesCommandList, kes_gpuptr_t dst, kes_gpuptr_t src, size_t size);
27
29
void (*fn_cmd_write_timestamp)(KesCommandList, kes_gpuptr_t addr);
28
30
void (*fn_cmd_signal_after)(KesCommandList, enum KesStage before, kes_gpuptr_t addr, uint64_t value, enum KesSignal);
29
31
void (*fn_cmd_wait_before)(KesCommandList, enum KesStage after, kes_gpuptr_t addr, uint64_t value, enum KesOp, enum KesHazardFlags, uint64_t mask);
+9
kestrel/include/kestrel/kestrel.h
+9
kestrel/include/kestrel/kestrel.h
···
243
243
void kes_cmd_memset(KesCommandList command_list, kes_gpuptr_t addr, size_t size, uint32_t value);
244
244
245
245
/**
246
+
* Record a memory copy command in the command list.
247
+
* @param command_list The command list to record the command in.
248
+
* @param dst The GPU address to copy memory to.
249
+
* @param src The GPU address to copy memory from.
250
+
* @param size The size of the copy in bytes.
251
+
*/
252
+
void kes_cmd_memcpy(KesCommandList command_list, kes_gpuptr_t dst, kes_gpuptr_t src, size_t size);
253
+
254
+
/**
246
255
* Record a timestamp write command in the command list. Writes a 64-bit unsigned timestamp in ticks to the provided address.
247
256
* @param command_list The command list to record the command in.
248
257
* @param addr The GPU address to write the timestamp to.
+8
kestrel/rt/api.cpp
+8
kestrel/rt/api.cpp
···
172
172
173
173
dev->fns.fn_cmd_memset(clhandle->cmdlist, addr, size, value);
174
174
}
175
+
176
+
API_EXPORT void kes_cmd_memcpy(KesCommandList pcl, kes_gpuptr_t dst, kes_gpuptr_t src, size_t size) {
177
+
auto *clhandle = reinterpret_cast<CommandListHandle *>(pcl);
178
+
auto *dev = clhandle->dev;
179
+
180
+
dev->fns.fn_cmd_memcpy(clhandle->cmdlist, dst, src, size);
181
+
}
182
+
175
183
API_EXPORT void kes_cmd_write_timestamp(KesCommandList pcl, kes_gpuptr_t addr) {
176
184
auto *clhandle = reinterpret_cast<CommandListHandle *>(pcl);
177
185
auto *dev = clhandle->dev;
+37
test/test/05_hello_transfer_copy/transfer_copy.cpp
+37
test/test/05_hello_transfer_copy/transfer_copy.cpp
···
1
+
#include <unistd.h>
2
+
#include <kestrel/kestrel.h>
3
+
4
+
#include <stdio.h>
5
+
6
+
int main(void) {
7
+
8
+
auto dev = kes_create();
9
+
10
+
std::size_t size = 10 * 1024 * 1024;
11
+
auto x = kes_malloc(dev, size, 4, KesMemoryDefault);
12
+
auto y = kes_malloc(dev, size, 4, KesMemoryDefault);
13
+
14
+
printf("x: %p (%p) (%llu bytes)\n", x.cpu, x.gpu, x.size);
15
+
printf("y: %p (%p) (%llu bytes)\n", y.cpu, y.gpu, y.size);
16
+
17
+
auto dma = kes_create_queue(dev, KesQueueTypeTransfer);
18
+
19
+
sleep(1);
20
+
21
+
auto l1 = kes_start_recording(dma);
22
+
kes_cmd_memset(l1, x.gpu, size, 2);
23
+
kes_cmd_memcpy(l1, y.gpu, x.gpu, size);
24
+
kes_submit(dma, l1);
25
+
26
+
// @todo: hacky bussy-wait
27
+
printf("x[0]: %u\n", ((uint32_t *)x.cpu)[0]);
28
+
printf("y[0]: %u\n", ((uint32_t *)y.cpu)[0]);
29
+
sleep(1);
30
+
printf("x[0]: %u\n", ((uint32_t *)x.cpu)[0]);
31
+
printf("y[0]: %u\n", ((uint32_t *)y.cpu)[0]);
32
+
33
+
kes_free(dev, &x);
34
+
kes_destroy(dev);
35
+
36
+
return 0;
37
+
}