适用于英特尔 CPU 的基本 C++ SIMD 类型 (sse、avx、avx2、avx512)
This library is header-only and is implemented according to which Intel ISA flags are enabled in the translation unit for which they are used (e.g. -mavx with gcc or clang).
(unofficial list of compilers, not all are tested)
The library is logically composed of 3 different components:
While there does not yet exist any true documentation, users are encouraged to
see what type aliases are defined in tsimd/detail/pack.h, as well as what
operators and functions are available in tsimd/detail/operators/ and
tsimd/detail/functions/ respectively. Generally speaking, each header
found in detail/ encapsulates exactly one type, operator, or function to
aide in discovery.
Consider the following function (kernel) taking values from two input arrays and storing in an output array.
// NOTE: n is the length of all 3 arrays
void saxpy(float a, int n, float x[], float y[], float out[])
{
for (int i = 0; i (&x[i]);
const vfloat yi = load(&y[i]);
const vfloat result = a * xi + yi; // same formula!
store(result, &out[i]);
}
}The advantage to this version (instead of using a specific SIMD width, say
vfloat4 or vfloat8) is that the kernel function will be "widened"
to the best available width based on how it gets compiled. In other words:
4-wide for SSE, 8-wide for AVX/AVX2, and 16-wide for AVX512.
暂无开放 Issues,或尚未同步最近议题。