TEST_P(ocl_engine_test, BasicInteropC) {
auto p = GetParam();
cl_device_id ocl_dev = (p.adev_kind == dev_kind::gpu) ?
gpu_ocl_dev :
(p.adev_kind == dev_kind::cpu) ? cpu_ocl_dev : nullptr;
cl_context ocl_ctx = (p.actx_kind == ctx_kind::gpu) ?
gpu_ocl_ctx :
(p.actx_kind == ctx_kind::cpu) ? cpu_ocl_ctx : nullptr;
SKIP_IF(p.adev_kind != dev_kind::null && !ocl_dev,
"Required OpenCL device not found.");
SKIP_IF(p.actx_kind != ctx_kind::null && !ocl_ctx,
"Required OpenCL context not found.");
SKIP_IF(cpu_ocl_dev == gpu_ocl_dev
&& (p.adev_kind == dev_kind::cpu
|| p.actx_kind == ctx_kind::cpu),
"OpenCL CPU-only device not found.");
mkldnn_engine_t eng;
mkldnn_status_t s
= mkldnn_engine_create_ocl(&eng, mkldnn_gpu, ocl_dev, ocl_ctx);
EXPECT_EQ(s, p.expected_status);
if (s == mkldnn_success) {
cl_device_id dev;
cl_context ctx;
MKLDNN_CHECK(mkldnn_engine_get_ocl_device(eng, &dev));
MKLDNN_CHECK(mkldnn_engine_get_ocl_context(eng, &ctx));
EXPECT_EQ(dev, ocl_dev);
EXPECT_EQ(ctx, ocl_ctx);
cl_uint ref_count;
OCL_CHECK(clGetContextInfo(ocl_ctx, CL_CONTEXT_REFERENCE_COUNT,
sizeof(ref_count), &ref_count, nullptr));
int i_ref_count = int(ref_count);
EXPECT_EQ(i_ref_count, 2);
MKLDNN_CHECK(mkldnn_engine_destroy(eng));
OCL_CHECK(clGetContextInfo(ocl_ctx, CL_CONTEXT_REFERENCE_COUNT,
sizeof(ref_count), &ref_count, nullptr));
i_ref_count = int(ref_count);
EXPECT_EQ(i_ref_count, 1);
}
}
TEST_P(ocl_engine_test, BasicInteropCpp) {
auto p = GetParam();
cl_device_id ocl_dev = (p.adev_kind == dev_kind::gpu) ?
gpu_ocl_dev :
(p.adev_kind == dev_kind::cpu) ? cpu_ocl_dev : nullptr;
cl_context ocl_ctx = (p.actx_kind == ctx_kind::gpu) ?
gpu_ocl_ctx :
(p.actx_kind == ctx_kind::cpu) ? cpu_ocl_ctx : nullptr;
SKIP_IF(p.adev_kind != dev_kind::null && !ocl_dev,
"Required OpenCL device not found.");
SKIP_IF(p.actx_kind != ctx_kind::null && !ocl_ctx,
"Required OpenCL context not found.");
SKIP_IF(cpu_ocl_dev == gpu_ocl_dev
&& (p.adev_kind == dev_kind::cpu
|| p.actx_kind == ctx_kind::cpu),
"OpenCL CPU-only device not found.");
catch_expected_failures(
[&]() {
{
engine eng(engine::kind::gpu, ocl_dev, ocl_ctx);
if (p.expected_status != mkldnn_success) {
FAIL() << "Success not expected";
}
cl_device_id dev = eng.get_ocl_device();
cl_context ctx = eng.get_ocl_context();
EXPECT_EQ(dev, ocl_dev);
EXPECT_EQ(ctx, ocl_ctx);
cl_uint ref_count;
OCL_CHECK(clGetContextInfo(ocl_ctx,
CL_CONTEXT_REFERENCE_COUNT, sizeof(ref_count),
&ref_count, nullptr));
int i_ref_count = int(ref_count);
EXPECT_EQ(i_ref_count, 2);
}
cl_uint ref_count;
OCL_CHECK(clGetContextInfo(ocl_ctx, CL_CONTEXT_REFERENCE_COUNT,
sizeof(ref_count), &ref_count, nullptr));
int i_ref_count = int(ref_count);
EXPECT_EQ(i_ref_count, 1);
},
p.expected_status != mkldnn_success, p.expected_status);
}
TEST_P(Int8ToInt32, UsingLoadParam) {
std::string arch = omrsysinfo_get_CPU_architecture();
SKIP_IF(OMRPORT_ARCH_S390 == arch || OMRPORT_ARCH_S390X == arch, KnownBug)
<< "The Z code generator incorrectly spills sub-integer types arguments (see issue #3525)";
auto param = TRTest::to_struct(GetParam());
char *inputTrees =
"(method return=Int32 args=[Int8]"
" (block"
" (ireturn"
" (b2i"
" (bload parm=0) ) ) ) )";
auto trees = parseString(inputTrees);
ASSERT_NOTNULL(trees);
Tril::DefaultCompiler compiler{trees};
ASSERT_EQ(0, compiler.compile()) << "Compilation failed unexpectedly\n" << "Input trees: " << inputTrees;
auto entry_point = compiler.getEntryPoint<int32_t (*)(int8_t)>();
volatile auto exp = param.oracle(param.value);
volatile auto act = entry_point(param.value);
ASSERT_EQ(exp, act);
}
TEST_P(engine_test_c, CreateWithBackend) {
mkldnn_engine_kind_t engine_kind = GetParam();
SKIP_IF(mkldnn_engine_get_count(engine_kind) == 0,
"Engine kind is not supported.");
auto backend_kinds = { mkldnn_backend_native, mkldnn_backend_ocl };
for (auto backend_kind : backend_kinds) {
if (engine_kind == mkldnn_cpu) {
if (backend_kind == mkldnn_backend_ocl)
continue;
} else if (engine_kind == mkldnn_gpu) {
if (backend_kind == mkldnn_backend_native)
continue;
}
mkldnn_engine_t engine;
MKLDNN_CHECK(mkldnn_engine_create_with_backend(
&engine, engine_kind, backend_kind, 0));
mkldnn_backend_kind_t queried_backend_kind;
MKLDNN_CHECK(
mkldnn_engine_get_backend_kind(engine, &queried_backend_kind));
EXPECT_EQ(queried_backend_kind, backend_kind);
MKLDNN_CHECK(mkldnn_engine_destroy(engine));
}
}
int main(void)
{
SKIP_IF(valgrind_is_caller());
ASSERT_TRUE(close(-1) == -1);
ASSERT_TRUE(errno == EBADF);
return EXIT_SUCCESS;
}
TEST_P(engine_test_cpp, GetBackend) {
engine::kind engine_kind = static_cast<engine::kind>(GetParam());
SKIP_IF(engine::get_count(engine_kind) == 0,
"Engine kind is not supported.");
engine eng(engine_kind, 0);
compare_backend(static_cast<mkldnn_engine_kind_t>(engine_kind),
static_cast<mkldnn_backend_kind_t>(eng.get_backend_kind()));
}
int cpu_event_pinned_vs_ebb(void)
{
union pipe read_pipe, write_pipe;
struct event event;
int cpu, rc;
pid_t pid;
SKIP_IF(!ebb_is_supported());
cpu = pick_online_cpu();
FAIL_IF(cpu < 0);
FAIL_IF(bind_to_cpu(cpu));
FAIL_IF(pipe(read_pipe.fds) == -1);
FAIL_IF(pipe(write_pipe.fds) == -1);
pid = fork();
if (pid == 0) {
/* NB order of pipes looks reversed */
exit(ebb_child(write_pipe, read_pipe));
}
/* We setup the cpu event first */
rc = setup_cpu_event(&event, cpu);
if (rc) {
kill_child_and_wait(pid);
return rc;
}
/* Signal the child to install its EBB event and wait */
if (sync_with_child(read_pipe, write_pipe))
/* If it fails, wait for it to exit */
goto wait;
/* Signal the child to run */
FAIL_IF(sync_with_child(read_pipe, write_pipe));
wait:
/* We expect it to fail to read the event */
FAIL_IF(wait_for_child(pid) != 2);
FAIL_IF(event_disable(&event));
FAIL_IF(event_read(&event));
event_report(&event);
/* The cpu event should have run */
FAIL_IF(event.result.value == 0);
FAIL_IF(event.result.enabled != event.result.running);
return 0;
}
runall_result test_main()
{
SKIP_IF(!is_gpu_hardware_available());
runall_result result;
result &= REPORT_RESULT((test_feature<int, 1>()));
result &= REPORT_RESULT((test_feature<unsigned int, 91>()));
result &= REPORT_RESULT((test_feature<float, 5>()));
result &= REPORT_RESULT((test_feature<double, 31>()));
return result;
}
int test_copy_unaligned(void)
{
/* Only run this test on a P9 or later */
SKIP_IF(!have_hwcap2(PPC_FEATURE2_ARCH_3_00));
/* Register our signal handler with SIGBUS */
setup_signal_handler();
/* +1 makes buf unaligned */
copy(cacheline_buf+1);
/* We should not get here */
return 1;
}
TEST_P(engine_test_c, GetBackend) {
mkldnn_engine_kind_t engine_kind = GetParam();
SKIP_IF(mkldnn_engine_get_count(engine_kind) == 0,
"Engine kind is not supported.");
mkldnn_engine_t engine;
MKLDNN_CHECK(mkldnn_engine_create(&engine, engine_kind, 0));
mkldnn_backend_kind_t backend_kind;
MKLDNN_CHECK(mkldnn_engine_get_backend_kind(engine, &backend_kind));
compare_backend(engine_kind, backend_kind);
MKLDNN_CHECK(mkldnn_engine_destroy(engine));
}
static int setup_cpu_event(struct event *event, int cpu)
{
event_init_named(event, 0x400FA, "PM_RUN_INST_CMPL");
event->attr.pinned = 1;
event->attr.exclude_kernel = 1;
event->attr.exclude_hv = 1;
event->attr.exclude_idle = 1;
SKIP_IF(require_paranoia_below(1));
FAIL_IF(event_open_with_cpu(event, cpu));
FAIL_IF(event_enable(event));
return 0;
}
runall_result test_main()
{
SKIP_IF(!is_gpu_hardware_available());
runall_result result;
result &= REPORT_RESULT((test_feature<int, 1, 1, 1>()));
result &= REPORT_RESULT((test_feature<int, 7, 3, 2>()));
result &= REPORT_RESULT((test_feature<int, 5, 3, 5>()));
result &= REPORT_RESULT((test_feature<unsigned, 3, 2, 1>()));
result &= REPORT_RESULT((test_feature<signed, 3, 5, 5>()));
result &= REPORT_RESULT((test_feature<float, 2, 3, 2>()));
result &= REPORT_RESULT((test_feature<float, 5, 1, 5>()));
result &= REPORT_RESULT((test_feature<double, 2, 7, 7>()));
return result;
}
int task_event_vs_ebb(void)
{
union pipe read_pipe, write_pipe;
struct event event;
pid_t pid;
int rc;
SKIP_IF(!ebb_is_supported());
FAIL_IF(pipe(read_pipe.fds) == -1);
FAIL_IF(pipe(write_pipe.fds) == -1);
pid = fork();
if (pid == 0) {
/* NB order of pipes looks reversed */
exit(ebb_child(write_pipe, read_pipe));
}
/* We setup the task event first */
rc = setup_child_event(&event, pid);
if (rc) {
kill_child_and_wait(pid);
return rc;
}
/* Signal the child to install its EBB event and wait */
if (sync_with_child(read_pipe, write_pipe))
/* If it fails, wait for it to exit */
goto wait;
/* Signal the child to run */
FAIL_IF(sync_with_child(read_pipe, write_pipe));
wait:
/* The EBB event should push the task event off so the child should succeed */
FAIL_IF(wait_for_child(pid));
FAIL_IF(event_disable(&event));
FAIL_IF(event_read(&event));
event_report(&event);
/* The task event may have run, or not so we can't assert anything about it */
return 0;
}
static int no_handler_test(void)
{
struct event event;
u64 val;
int i;
SKIP_IF(!ebb_is_supported());
event_init_named(&event, 0x1001e, "cycles");
event_leader_ebb_init(&event);
event.attr.exclude_kernel = 1;
event.attr.exclude_hv = 1;
event.attr.exclude_idle = 1;
FAIL_IF(event_open(&event));
FAIL_IF(ebb_event_enable(&event));
val = mfspr(SPRN_EBBHR);
FAIL_IF(val != 0);
/* Make sure it overflows quickly */
sample_period = 1000;
mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
/* Spin to make sure the event has time to overflow */
for (i = 0; i < 1000; i++)
mb();
dump_ebb_state();
/* We expect to see the PMU frozen & PMAO set */
val = mfspr(SPRN_MMCR0);
FAIL_IF(val != 0x0000000080000080);
event_close(&event);
dump_ebb_state();
/* The real test is that we never took an EBB at 0x0 */
return 0;
}
/*
* Test basic access to the EBB regs, they should be user accessible with no
* kernel interaction required.
*/
int reg_access(void)
{
uint64_t val, expected;
SKIP_IF(!ebb_is_supported());
expected = 0x8000000100000000ull;
mtspr(SPRN_BESCR, expected);
val = mfspr(SPRN_BESCR);
FAIL_IF(val != expected);
expected = 0x0000000001000000ull;
mtspr(SPRN_EBBHR, expected);
val = mfspr(SPRN_EBBHR);
FAIL_IF(val != expected);
return 0;
}
static int ipc_unmuxed(void)
{
int tests_done = 0;
#define DO_TEST(_name, _num) \
FAIL_IF(test_##_name()); \
tests_done++;
#include "ipc.h"
#undef DO_TEST
/*
* If we ran no tests then it means none of the syscall numbers were
* defined, possibly because we were built against old headers. But it
* means we didn't really test anything, so instead of passing mark it
* as a skip to give the user a clue.
*/
SKIP_IF(tests_done == 0);
return 0;
}
int tm_syscall(void)
{
unsigned count = 0;
struct timeval end, now;
SKIP_IF(!((long)get_auxv_entry(AT_HWCAP2)
& PPC_FEATURE2_HTM_NOSC));
setbuf(stdout, NULL);
printf("Testing transactional syscalls for %d seconds...\n", TEST_DURATION);
gettimeofday(&end, NULL);
now.tv_sec = TEST_DURATION;
now.tv_usec = 0;
timeradd(&end, &now, &end);
for (count = 0; timercmp(&now, &end, <); count++) {
/*
* Test a syscall within a suspended transaction and verify
* that it succeeds.
*/
FAIL_IF(getppid_tm(true) == -1); /* Should succeed. */
/*
* Test a syscall within an active transaction and verify that
* it fails with the correct failure code.
*/
FAIL_IF(getppid_tm(false) != -1); /* Should fail... */
FAIL_IF(!failure_is_persistent()); /* ...persistently... */
FAIL_IF(!failure_is_syscall()); /* ...with code syscall. */
gettimeofday(&now, 0);
}
printf("%d active and suspended transactions behaved correctly.\n", count);
printf("(There were %d transaction retries.)\n", retries);
return 0;
}
static int per_event_excludes(void)
{
struct event *e, events[4];
char *platform;
int i;
platform = (char *)get_auxv_entry(AT_BASE_PLATFORM);
FAIL_IF(!platform);
SKIP_IF(strcmp(platform, "power8") != 0);
/*
* We need to create the events disabled, otherwise the running/enabled
* counts don't match up.
*/
e = &events[0];
event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
PERF_TYPE_HARDWARE, "instructions");
e->attr.disabled = 1;
e = &events[1];
event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
PERF_TYPE_HARDWARE, "instructions(k)");
e->attr.disabled = 1;
e->attr.exclude_user = 1;
e->attr.exclude_hv = 1;
e = &events[2];
event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
PERF_TYPE_HARDWARE, "instructions(h)");
e->attr.disabled = 1;
e->attr.exclude_user = 1;
e->attr.exclude_kernel = 1;
e = &events[3];
event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
PERF_TYPE_HARDWARE, "instructions(u)");
e->attr.disabled = 1;
e->attr.exclude_hv = 1;
e->attr.exclude_kernel = 1;
FAIL_IF(event_open(&events[0]));
/*
* The open here will fail if we don't have per event exclude support,
* because the second event has an incompatible set of exclude settings
* and we're asking for the events to be in a group.
*/
for (i = 1; i < 4; i++)
FAIL_IF(event_open_with_group(&events[i], events[0].fd));
/*
* Even though the above will fail without per-event excludes we keep
* testing in order to be thorough.
*/
prctl(PR_TASK_PERF_EVENTS_ENABLE);
/* Spin for a while */
for (i = 0; i < INT_MAX; i++)
asm volatile("" : : : "memory");
prctl(PR_TASK_PERF_EVENTS_DISABLE);
for (i = 0; i < 4; i++) {
FAIL_IF(event_read(&events[i]));
event_report(&events[i]);
}
/*
* We should see that all events have enabled == running. That
* shows that they were all on the PMU at once.
*/
for (i = 0; i < 4; i++)
FAIL_IF(events[i].result.running != events[i].result.enabled);
/*
* We can also check that the result for instructions is >= all the
* other counts. That's because it is counting all instructions while
* the others are counting a subset.
*/
for (i = 1; i < 4; i++)
FAIL_IF(events[0].result.value < events[i].result.value);
for (i = 0; i < 4; i++)
event_close(&events[i]);
return 0;
}
