A Modern GPGPU API & wip linux RDNA2+ Driver
rdna
driver
linux
gpu
amdgpu: cleanup
+30
-12
+18
-3
drivers/amdgpu/cmds.cpp
+18
-3
drivers/amdgpu/cmds.cpp
···
296
296
void init_compute_shader_config(DeviceImpl *dev, Shader &shader) {
297
297
298
298
// @todo: ultra temporary.
299
-
auto x = amdgpu_malloc(dev, 1024, 16, KesMemoryDefault);
300
-
*((uint32_t *)x.cpu) = 0xBF810000; // s_endpgm
299
+
auto x = amdgpu_malloc(dev, 1024, 256, KesMemoryDefault);
300
+
uint32_t *pgm = (uint32_t *)x.cpu;
301
+
auto pgmidx = 0;
302
+
// pgm[pgmidx++] = 0x24020402; // v_lshlrev_b32 v1, 2, v0
303
+
// pgm[pgmidx++] = 0x4A020300; // v_add_co_u32 v1, vcc_lo, s0, v1
304
+
// pgm[pgmidx++] = 0x4A040201; // v_add_co_ci_u32 v2, vcc_lo, s1, 0, vcc_lo
305
+
// pgm[pgmidx++] = 0xDC500000; // global_store_dword v[1:2], v0, off
306
+
// pgm[pgmidx++] = 0x00000001;
307
+
pgm[pgmidx++] = 0xBF810000; // s_endpgm
308
+
309
+
// (RDNA ISA Ref. 2.5)
310
+
for (auto i = 0; i < 64; ++i) {
311
+
pgm[pgmidx++] = 0xBF9F0000; // s_code_end
312
+
}
313
+
314
+
log("shader code: {} {}", (void *)x.cpu, (void *)x.gpu);
301
315
302
316
// @todo: temporary
303
317
auto wave_size = 32;
···
334
348
335
349
shader.config.pgm_rsrc1 =
336
350
S_00B848_VGPRS((num_vgprs - 1) / (wave_size == 32 ? 8 : 4))
337
-
| S_00B848_DX10_CLAMP(dx10_clamp);
351
+
| S_00B848_DX10_CLAMP(dx10_clamp)
352
+
| S_00B128_MEM_ORDERED(true); //always true for gfx10.3
338
353
339
354
shader.config.pgm_rsrc2 =
340
355
S_00B84C_USER_SGPR(shader.config.user_sgpr_count)
+12
-9
test/examples/07_hello_dispatch/hello_dispatch.cpp
+12
-9
test/examples/07_hello_dispatch/hello_dispatch.cpp
···
4
4
#include <stdio.h>
5
5
6
6
struct DispatchArguments {
7
-
uint64_t va;
8
-
uint32_t size;
7
+
uint64_t buffer;
9
8
};
10
9
11
10
int main(void) {
12
-
13
-
auto size = 10 * 1024 * 1024;
14
-
15
11
auto dev = kes_create();
16
12
17
-
auto x = kes_malloc(dev, size, 4, KesMemoryDefault);
13
+
auto x = kes_malloc(dev, 1024, 4, KesMemoryDefault);
18
14
auto y = kes_malloc(dev, sizeof(DispatchArguments), 8, KesMemoryDefault);
15
+
16
+
printf("x: %p %p\n", (void *)x.cpu, (void *)x.gpu);
17
+
printf("y: %p %p\n", (void *)y.cpu, (void *)y.gpu);
19
18
20
19
DispatchArguments *args = (DispatchArguments *)y.cpu;
21
-
args->va = x.gpu;
22
-
args->size = size;
20
+
args->buffer = x.gpu;
23
21
24
22
auto compute = kes_create_queue(dev, KesQueueTypeCompute);
25
23
26
24
auto cl = kes_start_recording(compute);
27
25
{
28
-
kes_cmd_dispatch(cl, y.gpu, 128, 1, 1);
26
+
kes_cmd_dispatch(cl, y.gpu, 32, 1, 1);
29
27
}
30
28
31
29
kes_submit(compute, cl);
32
30
33
31
sleep(1);
32
+
33
+
printf("x[0]: %u\n", ((uint32_t *)x.cpu)[0]);
34
+
printf("x[1]: %u\n", ((uint32_t *)x.cpu)[1]);
35
+
printf("x[2]: %u\n", ((uint32_t *)x.cpu)[2]);
36
+
printf("x[3]: %u\n", ((uint32_t *)x.cpu)[3]);
34
37
35
38
kes_free(dev, &x);
36
39
kes_destroy(dev);