1 /* 2 * Fused Multiply Add (Single) 3 * 4 * Copyright (c) 2019, 2024 Linaro 5 * 6 * SPDX-License-Identifier: GPL-2.0-or-later 7 */ 8 9 #include <stdio.h> 10 #include <stdlib.h> 11 #include <math.h> 12 #include <float.h> 13 #include <fenv.h> 14 15 #include "float_helpers.h" 16 17 #define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0])) 18 19 typedef struct { 20 int flag; 21 char *desc; 22 } float_mapping; 23 24 float_mapping round_flags[] = { 25 { FE_TONEAREST, "to nearest" }, 26 #ifdef FE_UPWARD 27 { FE_UPWARD, "upwards" }, 28 #endif 29 #ifdef FE_DOWNWARD 30 { FE_DOWNWARD, "downwards" }, 31 #endif 32 #ifdef FE_TOWARDZERO 33 { FE_TOWARDZERO, "to zero" } 34 #endif 35 }; 36 37 38 static void print_inputs(float a, float b, float c) 39 { 40 char *a_fmt, *b_fmt, *c_fmt; 41 42 a_fmt = fmt_f32(a); 43 b_fmt = fmt_f32(b); 44 c_fmt = fmt_f32(c); 45 46 printf("op : %s * %s + %s\n", a_fmt, b_fmt, c_fmt); 47 48 free(a_fmt); 49 free(b_fmt); 50 free(c_fmt); 51 } 52 53 static void print_result(float r, int j, int k) 54 { 55 char *r_fmt, *flag_fmt; 56 57 flag_fmt = fmt_flags(); 58 r_fmt = fmt_f32(r); 59 60 printf("res: %s flags=%s (%d/%d)\n", r_fmt, flag_fmt, j, k); 61 62 free(r_fmt); 63 free(flag_fmt); 64 } 65 66 static void do_madds(float a, float b, float c, int j, int k) 67 { 68 float r; 69 70 print_inputs(a, b, c); 71 72 feclearexcept(FE_ALL_EXCEPT); 73 r = __builtin_fmaf(a, b, c); 74 75 print_result(r, j, k); 76 } 77 78 int main(int argc, char *argv[argc]) 79 { 80 int i, j, k, nums = get_num_f32(); 81 float a, b, c; 82 83 for (i = 0; i < ARRAY_SIZE(round_flags); ++i) { 84 if (fesetround(round_flags[i].flag) != 0) { 85 printf("### Rounding %s skipped\n", round_flags[i].desc); 86 continue; 87 } 88 printf("### Rounding %s\n", round_flags[i].desc); 89 for (j = 0; j < nums; j++) { 90 for (k = 0; k < 3; k++) { 91 a = get_f32(j + ((k)%3)); 92 b = get_f32(j + ((k+1)%3)); 93 c = get_f32(j + ((k+2)%3)); 94 do_madds(a, b, c, j, k); 95 } 96 } 97 98 /* From https://bugs.launchpad.net/qemu/+bug/1841491 */ 99 printf("# LP184149\n"); 100 do_madds(0x1.ffffffffffffcp-1022, 0x1.0000000000001p-1, 0x0.0000000000001p-1022, j, 0); 101 do_madds(0x8p-152, 0x8p-152, 0x8p-152, j+1, 0); 102 } 103 104 return 0; 105 } 106