def test_metadata_run_traces(profile, format):
data = profile("""
__global__ void kernel(int* p) {
*p = threadIdx.x;
}
__global__ void kernel2(int* p) {
*p = threadIdx.x;
}
int main()
{
int* dptr;
cudaMalloc(&dptr, sizeof(int));
kernel<<<1, 1>>>(dptr);
kernel<<<1, 1>>>(dptr);
kernel2<<<1, 1>>>(dptr);
cudaFree(dptr);
return 0;
}
""",
format=format)
run = data[run_file()]
traces = run["traces"]
assert "kernel-0.trace." + format in traces
assert "kernel-1.trace." + format in traces
assert "kernel2-0.trace." + format in traces
def test_general_no_include(profile, format):
data = profile("""
__global__ void kernel(int* p) {
*p = 5;
}
int main() {
int* dptr;
cudaMalloc(&dptr, sizeof(int));
kernel<<<1, 1>>>(dptr);
cudaFree(dptr);
return 0;
}
""",
format=format,
add_include=False)
assert run_file() in data
assert len(data) == 1
def test_general_emit_nothing(profile, format):
data = profile("", with_main=True)
assert run_file() in data
assert len(data) == 1
def test_parameters_compression_run(profile, format):
uncompressed = profile("", with_main=True, format=format)
compressed = profile("", with_main=True, format=format, compress=True)
assert not uncompressed[run_file()]["compress"]
assert compressed[run_file()]["compress"]
def test_metadata_run_time(profile):
data = profile("__global__ void kernel() {}", with_main=True)
run = data[run_file()]
assert run["type"] == "run"
assert run["end"] >= run["start"]