Visible to Intel only — GUID: GUID-89A2552B-6B48-4FB2-B870-F7427CC7DECE
Visible to Intel only — GUID: GUID-89A2552B-6B48-4FB2-B870-F7427CC7DECE
DPCT1040
Message
Use sycl::stream instead of printf if your code is used on the device.
Detailed Help
If the printf statement is used on the host and the device in your original code, it does not change. To create output in DPC++, sycl::stream must be used on the device and printf must be used on the host.
Suggestions to Fix
If the printf statement is only used from the host, do not change your code.
If the printf statement is only used from the device, use sycl::stream instead of printf.
For example, this original CUDA* code:
__host__ __device__ void hd() {
printf("Hello!\n");
}
__global__ void k() {
hd();
}
void foo() {
hd();
k<<<1, 1>>>();
}
results in the following migrated SYCL* code:
void hd() {
/*
DPCT1040:0: Use sycl::stream instead of printf if your code is used on the
device.
*/
printf("Hello!\n");
}
void k() {
hd();
}
void foo() {
hd();
dpct::get_default_queue().parallel_for(
sycl::nd_range<3>(sycl::range<3>(1, 1, 1), sycl::range<3>(1, 1, 1)),
[=](sycl::nd_item<3> item_ct1) {
k();
});
}
which is rewritten to:
void hd_host() {
printf("Hello!\n");
}
void hd_device(const sycl::stream &stream) {
stream << "Hello!\n";
}
void k(const sycl::stream &stream) {
hd_device(stream);
}
void foo() {
hd_host();
dpct::get_default_queue().submit([&](sycl::handler &cgh) {
sycl::stream stream(64 * 1024 /*totalBufferSize*/, 80 /*workItemBufferSize*/, cgh);
cgh.parallel_for(
sycl::nd_range<3>(sycl::range<3>(1, 1, 1), sycl::range<3>(1, 1, 1)),
[=](sycl::nd_item<3> item_ct1) {
k(stream);
});
});
}