Visible to Intel only — GUID: GUID-F030395F-E0F2-4952-98E4-FAD78D1AAE8F
Visible to Intel only — GUID: GUID-F030395F-E0F2-4952-98E4-FAD78D1AAE8F
DPCT1038
Message
When the kernel function name is used as a macro argument, the migration result may be incorrect. You need to verify the definition of the macro.
Detailed Help
When a kernel function call is used in a macro definition and the function name is passed as a macro argument, the tool cannot determine the type of function parameters. This may lead to incorrectly generated DPC++ code.
Suggestions to Fix
Review the kernel call inside the macro definition and adjust it manually, if needed.
For example, this original CUDA* code:
__global__ void kernel(int c, int d) {}
void foo() {
#define KERNEL_LAUNCH(NAME, a, b, c, d) NAME<<<a, b, 0>>>(c, d);
KERNEL_LAUNCH(kernel, 3, 2, 1, 0)
}
results in the following migrated SYCL* code:
void kernel(int c, int d) {}
void foo() {
#define KERNEL_LAUNCH(NAME, a, b, c, d) \
dpct::get_default_queue().submit([&](sycl::handler &cgh) { \
auto c_ct0 = c; \
auto d_ct1 = d; \
\
cgh.parallel_for( \
sycl::nd_range<3>(sycl::range<3>(1, 1, a) * sycl::range<3>(1, 1, b), \
sycl::range<3>(1, 1, b)), \
[=](sycl::nd_item<3> item_ct1) { kernel(c_ct0, d_ct1); }); \
});
/*
DPCT1038:0: When the kernel function name is used as a macro argument, the
migration result may be incorrect. You need to verify the definition of the
macro.
*/
KERNEL_LAUNCH(kernel, 3, 2, 1, 0)
}
which is rewritten to:
void kernel(int c, int d) {}
void foo() {
#define KERNEL_LAUNCH(NAME, a, b, c, d) \
dpct::get_default_queue().submit([&](sycl::handler &cgh) { \
auto c_ct0 = c; \
auto d_ct1 = d; \
\
cgh.parallel_for( \
sycl::nd_range<3>(sycl::range<3>(1, 1, a) * sycl::range<3>(1, 1, b), \
sycl::range<3>(1, 1, b)), \
[=](sycl::nd_item<3> item_ct1) { NAME(c_ct0, d_ct1); }); \
});
KERNEL_LAUNCH(kernel, 3, 2, 1, 0)
}