Test program to list inputs for which vfrec7.v approximation is inexact in BFloat16

vfrec7_bf16.c

/** Determining when RVV 1.0 vfrev7.v SEW=32 result can not be converted
 *  exactly to a BF16 value. */
#include <stdio.h>
#include <stdint.h>
#include <inttypes.h>


typedef union {
    float f;
    uint32_t u;
} fu_t;

int main(void) {

    for (int i = 0; i < 256; ++i) {
        uint32_t input = 42, fdst = 17;
        __asm__ volatile(
            "slli a0, %[i], 16\n"
            "li a1, 0x7e800000\n"
            "add %[input], a0, a1\n"
            "vsetivli zero, 1, e32, m1, ta, ma\n"
            "vmv.v.x v0, %[input]\n"
            "vfrec7.v v0, v0\n"
            "vmv.x.s %[fdst], v0"
             : [input]"=r"(input), [fdst]"=r"(fdst): [i]"r"(i)
             : "a0", "a1", "v0", "memory"
        );
        fu_t f;
        f.u = input;
        printf("%d %a 1/%"PRIx32" = %"PRIx32" BF16 underflow %d\n", i, (double) f.f, input, fdst, (fdst & 0xffff) != 0);

    }

    return 0;

}