A Modern GPGPU API & wip linux RDNA2+ Driver
rdna
driver
linux
gpu
test: added unit tests (broken)
+60
+60
test/unit/test.cpp
+60
test/unit/test.cpp
···
1
+
#include "test_utils.h"
2
+
#include <catch2/catch_test_macros.hpp>
3
+
#include <catch2/generators/catch_generators.hpp>
4
+
#include <cstdlib>
5
+
6
+
TEST_CASE("Device creation and alloc") {
7
+
DeviceGuard device;
8
+
9
+
auto x = kes_malloc(device.dev, 256, 4, KesMemoryDefault);
10
+
11
+
REQUIRE(x.cpu != 0);
12
+
REQUIRE(x.gpu != 0);
13
+
}
14
+
15
+
/*
16
+
TEST_CASE("Cross-queue synchronization", "[sync][queue]") {
17
+
DeviceGuard device;
18
+
19
+
// Test parameters
20
+
auto size = GENERATE(1024, 1024*1024, 10*1024*1024);
21
+
22
+
SECTION("Transfer -> Compute sync with signal") {
23
+
auto buffer = device.alloc(size);
24
+
auto signal_mem = device.alloc(4);
25
+
TimestampReader timestamps(device.dev, 4);
26
+
27
+
auto dma = kes_create_queue(device.dev, KesQueueTypeTransfer);
28
+
auto compute = kes_create_queue(device.dev, KesQueueTypeCompute);
29
+
30
+
// DMA: memset + signal
31
+
auto l1 = kes_start_recording(dma);
32
+
kes_cmd_write_timestamp(l1, timestamps.gpu_ptr(0));
33
+
kes_cmd_memset(l1, buffer.gpu, size, 0x42);
34
+
kes_cmd_write_timestamp(l1, timestamps.gpu_ptr(1));
35
+
kes_cmd_signal_after(l1, KesStageTransfer, signal_mem.gpu,
36
+
0x13, KesSignalAtomicSet);
37
+
38
+
// Compute: wait + timestamp
39
+
auto l2 = kes_start_recording(compute);
40
+
kes_cmd_write_timestamp(l2, timestamps.gpu_ptr(2));
41
+
kes_cmd_wait_before(l2, KesStageTransfer, signal_mem.gpu,
42
+
0x13, KesOpEqual, KesHazardFlagsNone, ~0);
43
+
kes_cmd_write_timestamp(l2, timestamps.gpu_ptr(3));
44
+
45
+
kes_submit(dma, l1);
46
+
kes_submit(compute, l2);
47
+
48
+
// Wait for completion
49
+
busy_wait_for_value(signal_mem.cpu, 0x13);
50
+
51
+
// Verify memory was set
52
+
// REQUIRE(((uint32_t *)buffer.cpu)[0] == 0x42);
53
+
// REQUIRE(((uint32_t *)buffer.cpu)[size-1] == 0x42);
54
+
55
+
// Verify timestamp ordering: t0 <= t1 <= t3, t2 <= t3
56
+
timestamps.check_ordering({0, 1, 3});
57
+
timestamps.check_ordering({2, 3});
58
+
}
59
+
}
60
+
*/