numerics-astepin

Hello @jinge90 ```cuda #include #include #include #include extern "C" __device__ double __nv_rcp64h(double a); __device__ void calculate(const double* a, double* y) { y[0] = __nv_rcp64h(a[0]); } __global__ void test_kernel(size_t n, const...

@jinge90 I believe it flushes denormals to zero in source and destination and utilizes a slightly different table. 1-bit differences are unavoidable: ``` 0x1.eeeeeeeeeeeeep-1022 => 0x1.08d3e00000000p+1021 vs 0x1.eeeeeeeeeeeeep-1022 => 0x1.08d3d00000000p+1021...

@jinge90 rcp64h provides initial value for division algorithm to work. Sometimes such algorithms are implemented as table-lookup. ``` 0x1.08d3e00000000p-1 real math infinite prec. machine result 1.0 / 5.0 = 1.999999999999......p-3...

I suggest to even out with FTZ/DAZ behavior. I wrote this quickly to test new behavior. ```c++ void emulate(const double* a, double* y) { uint64_t xa = 0; uint64_t xy...