A Modern GPGPU API & wip linux RDNA2+ Driver
rdna
driver
linux
gpu
examples: move into new folder
+120
-12
+48
-12
test/CMakeLists.txt
+48
-12
test/CMakeLists.txt
···
1
+
include(FetchContent)
2
+
3
+
FetchContent_Declare(
4
+
Catch2
5
+
GIT_REPOSITORY https://github.com/catchorg/Catch2.git
6
+
GIT_TAG v3.5.1
7
+
)
8
+
FetchContent_MakeAvailable(Catch2)
1
9
2
-
file(GLOB TEST_DIRS CONFIGURE_DEPENDS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR}/test ${CMAKE_CURRENT_SOURCE_DIR}/test/*)
10
+
add_library(kestrel_test_common STATIC
11
+
common/test_utils.cpp
12
+
)
13
+
target_link_libraries(kestrel_test_common PUBLIC kestrel Catch2::Catch2)
14
+
target_include_directories(kestrel_test_common PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/common)
3
15
4
-
foreach(test_dir ${TEST_DIRS})
16
+
file(GLOB UNIT_TEST_SOURCES CONFIGURE_DEPENDS "unit/*.cpp")
17
+
18
+
if(UNIT_TEST_SOURCES)
19
+
add_executable(kestrel_unit_tests ${UNIT_TEST_SOURCES})
20
+
target_link_libraries(kestrel_unit_tests PRIVATE kestrel_test_common Catch2::Catch2WithMain)
21
+
add_sanitizers(kestrel_unit_tests)
22
+
23
+
# Register with CTest
24
+
add_test(NAME unit_tests COMMAND kestrel_unit_tests)
25
+
26
+
# These tests need GPU access
27
+
set_tests_properties(unit_tests PROPERTIES
28
+
LABELS "gpu"
29
+
TIMEOUT 300
30
+
)
31
+
endif()
32
+
33
+
include(CTest)
34
+
include(Catch)
35
+
36
+
if(TARGET kestrel_unit_tests)
37
+
catch_discover_tests(kestrel_unit_tests)
38
+
endif()
39
+
40
+
file(GLOB EXAMPLE_DIRS CONFIGURE_DEPENDS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR}/examples ${CMAKE_CURRENT_SOURCE_DIR}/examples/*)
41
+
foreach(example_dir ${EXAMPLE_DIRS})
5
42
# Ensure we are looking at a directory, not a random file
6
-
if(IS_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/test/${test_dir})
43
+
if(IS_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/examples/${example_dir})
7
44
8
45
# 2. Define the executable name based on the folder name
9
-
set(test_name "test_${test_dir}")
46
+
set(example_name "example_${example_dir}")
10
47
11
-
# 3. Find all .cpp files within that specific test folder
12
-
file(GLOB_RECURSE TEST_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/test/${test_dir}/*.cpp")
48
+
# 3. Find all .cpp files within that specific folder
49
+
file(GLOB_RECURSE EXAMPLE_SOURCES "${CMAKE_CURRENT_SOURCE_DIR}/examples/${example_dir}/*.cpp")
13
50
14
-
if(TEST_SOURCES)
15
-
add_executable(${test_name} ${TEST_SOURCES})
51
+
if(EXAMPLE_SOURCES)
52
+
add_executable(${example_name} ${EXAMPLE_SOURCES})
16
53
17
-
# 4. Link your library (libvektor)
18
-
target_link_libraries(${test_name} PRIVATE kestrel)
54
+
target_link_libraries(${example_name} PRIVATE kestrel)
19
55
20
-
add_sanitizers(${test_name})
56
+
add_sanitizers(${example_name})
21
57
22
-
message(STATUS "Added test: ${test_name}")
58
+
message(STATUS "Added example: ${example_name}")
23
59
endif()
24
60
endif()
25
61
endforeach()
+66
test/common/test_utils.h
+66
test/common/test_utils.h
···
1
+
#pragma once
2
+
3
+
#include <kestrel/kestrel.h>
4
+
#include <catch2/catch_test_macros.hpp>
5
+
#include <catch2/generators/catch_generators.hpp>
6
+
#include <vector>
7
+
#include <chrono>
8
+
#include <thread>
9
+
10
+
struct DeviceGuard {
11
+
KesDevice dev;
12
+
13
+
DeviceGuard() : dev(kes_create()) {
14
+
REQUIRE(dev != nullptr);
15
+
}
16
+
17
+
~DeviceGuard() {
18
+
if (dev) kes_destroy(dev);
19
+
}
20
+
21
+
DeviceGuard(const DeviceGuard&) = delete;
22
+
DeviceGuard& operator=(const DeviceGuard&) = delete;
23
+
24
+
KesAllocation alloc(size_t size, size_t align = 4, KesMemory type = KesMemoryDefault) {
25
+
return kes_malloc(dev, size, align, type);
26
+
}
27
+
};
28
+
29
+
struct TimestampReader {
30
+
KesAllocation alloc;
31
+
32
+
TimestampReader(KesDevice dev, size_t count = 8) {
33
+
alloc = kes_malloc(dev, sizeof(uint64_t) * count, 8, KesMemoryDefault);
34
+
}
35
+
36
+
uint64_t get(size_t index) const {
37
+
return ((uint64_t *)alloc.cpu)[index];
38
+
}
39
+
40
+
kes_gpuptr_t gpu_ptr(size_t index = 0) const {
41
+
return alloc.gpu + index * sizeof(uint64_t);
42
+
}
43
+
44
+
// Check ordering: timestamps should be non-decreasing
45
+
void check_ordering(std::vector<size_t> indices) const {
46
+
for (size_t i = 1; i < indices.size(); i++) {
47
+
uint64_t prev = get(indices[i-1]);
48
+
uint64_t curr = get(indices[i]);
49
+
INFO("Timestamp[" << indices[i-1] << "] = " << prev);
50
+
INFO("Timestamp[" << indices[i] << "] = " << curr);
51
+
REQUIRE(prev <= curr);
52
+
}
53
+
}
54
+
};
55
+
56
+
// CPU-side wait for GPU completion (for testing)
57
+
inline void busy_wait_for_value(void* ptr, uint64_t expected,
58
+
std::chrono::milliseconds timeout = std::chrono::seconds(5)) {
59
+
auto start = std::chrono::steady_clock::now();
60
+
while (*static_cast<volatile uint64_t*>(ptr) != expected) {
61
+
if (std::chrono::steady_clock::now() - start > timeout) {
62
+
FAIL("Timeout waiting for GPU value");
63
+
}
64
+
std::this_thread::sleep_for(std::chrono::microseconds(100));
65
+
}
66
+
}
test/test/01_hello_malloc/hello_malloc.cpp
test/examples/01_hello_malloc/hello_malloc.cpp
test/test/01_hello_malloc/hello_malloc.cpp
test/examples/01_hello_malloc/hello_malloc.cpp
test/test/02_hello_queue/hello_queue.cpp
test/examples/02_hello_queue/hello_queue.cpp
test/test/02_hello_queue/hello_queue.cpp
test/examples/02_hello_queue/hello_queue.cpp
test/test/03_hello_2queue/hello_2queue.cpp
test/examples/03_hello_2queue/hello_2queue.cpp
test/test/03_hello_2queue/hello_2queue.cpp
test/examples/03_hello_2queue/hello_2queue.cpp
test/test/04_hello_timestamp/hello_timestamp.cpp
test/examples/04_hello_timestamp/hello_timestamp.cpp
test/test/04_hello_timestamp/hello_timestamp.cpp
test/examples/04_hello_timestamp/hello_timestamp.cpp
test/test/05_hello_transfer_copy/transfer_copy.cpp
test/examples/05_hello_transfer_copy/transfer_copy.cpp
test/test/05_hello_transfer_copy/transfer_copy.cpp
test/examples/05_hello_transfer_copy/transfer_copy.cpp
+4
test/test/06_hello_sync/hello_sync.cpp
test/examples/06_hello_sync/hello_sync.cpp
+4
test/test/06_hello_sync/hello_sync.cpp
test/examples/06_hello_sync/hello_sync.cpp
···
11
11
auto x = kes_malloc(dev, size, 4, KesMemoryDefault);
12
12
auto y = kes_malloc(dev, 8, 4, KesMemoryDefault);
13
13
auto ts = kes_malloc(dev, 8 * 4, 4, KesMemoryDefault);
14
+
sleep(1);
14
15
15
16
printf("x: %p (%p) (%llu bytes)\n", x.cpu, x.gpu, x.size);
16
17
printf("y: %p (%p) (%llu bytes)\n", y.cpu, y.gpu, y.size);
17
18
18
19
auto dma = kes_create_queue(dev, KesQueueTypeTransfer);
19
20
auto compute = kes_create_queue(dev, KesQueueTypeCompute);
21
+
sleep(1);
20
22
21
23
auto l1 = kes_start_recording(dma);
22
24
{
···
44
46
printf("x[0]: %u\n", ((uint32_t *)x.cpu)[0]);
45
47
sleep(1);
46
48
printf("x[0]: %u\n", ((uint32_t *)x.cpu)[0]);
49
+
50
+
// in this case, synch should ensure that t0 <= t1, t1 <= t3, t2 <= t3
47
51
48
52
printf("\n");
49
53
printf("ts0: %lu\n", ((uint64_t *)ts.cpu)[0]);