oneAPI provides an Explicit SIMD SYCL extension (ESIMD) for lower-level Intel GPU programming.
ESIMD provides APIs that are similar to Intel's GPU Instruction Set Architecture (ISA), but it enables you to write explicitly vectorized device code. This explicit enabling gives you more control over the generated code and allows you to depend less on compiler optimizations.
The specification, API reference, and working code examples are available on GitHub*.
ESIMD kernels and functions always require a subgroup size of one, which means that the compiler does not provide vectorization across work items in a subgroup. Instead, you must explicitly express the vectorization in your code. Below is an example that adds the elements of two arrays and writes the results to the third:
float *A = malloc_shared<float>(Size, q);
float *B = malloc_shared<float>(Size, q);
float *C = malloc_shared<float>(Size, q);
for (unsigned i = 0; i != Size; i++) {
A[i] = B[i] = i;
}
q.submit([&](handler &cgh) {
cgh.parallel_for<class Test>(
Size / VL, [=](id<1> i)[[intel::sycl_explicit_simd]] {
auto offset = i * VL;
// pointer arithmetic, so offset is in elements:
simd<float, VL> va(A + offset);
simd<float, VL> vb(B + offset);
simd<float, VL> vc = va + vb;
vc.copy_to(C + offset);
});
}).wait_and_throw();
In the example above, the lambda function passed to the parallel_for is marked with a special attribute: [[intel::sycl_explicit_simd]]. This attribute tells the compiler that the kernel is ESIMD-based and ESIMD APIs can be used inside it. Here the simd objects and copy_to intrinsics are used. They are available only in the ESIMD extension.
Fully runnable code samples can be found on GitHub.
Code that uses the ESIMD extension can be compiled and run using the same commands as you would with standard SYCL:
To compile using the open-source oneAPI DPC++ Compiler:
$ clang++ -fsycl vadd_usm.cpp
To compile using an Intel® oneAPI Toolkit:
$ dpcpp vadd_usm.cpp
To run on an Intel specific GPU device, through the oneAPI Level Zero (Level Zero) backend:
$ SYCL_DEVICE_FILTER=level_zero:gpu ./a.out
The resulting executable ($./a.out) can be run only on Intel GPU hardware, such as Intel® UHD Graphics 600 or later. The DPC++ runtime automatically recognizes ESIMD kernels and dispatches their execution, so no additional setup is needed. Both Linux* and Windows* platforms are supported, including OpenCL™ and Level Zero backends.
This section contains lists of the main restrictions that apply when using the ESIMD extension.