F16C test code from gtk4

readme.md

In Fedora 36 on Parallels on a Core i9-based Mac, GTK4 apps are failing on an F16C extension instruction for converting between 32-bit and 16-bit floats for graphics drawing. This extracted code fails in the Parallels VM but works on the native MacOS. Linux's /proc/cpuinfo lists the f16c extension as present.

Expected output:

% make && ./borf
make: `borf' is up to date.
orig: -1.000000 128.500000 15000.000000 150000.000000
made small...
half as ints: bc00 5804 7353 7c00
made big again...
post: -1.000000 128.500000 15000.000000 inf

borf.c

/* fp16i.c
 *
 * Copyright 2021 Red Hat, Inc.
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this program.  If not, see <http://www.gnu.org/licenses/>.
 *
 * SPDX-License-Identifier: LGPL-2.1-or-later
 */

#define HAVE_F16C

#include <stdlib.h>
#include <stdio.h>
#include <inttypes.h>

#include <immintrin.h>

#if defined(_MSC_VER) && !defined(__clang__)
#define CAST_M128I_P(a) (__m128i const *) a
#else
#define CAST_M128I_P(a) (__m128i_u const *) a
#endif
void
float_to_half4_f16c (const float f[4],
                     uint16_t     h[4])
{
  __m128 s = _mm_loadu_ps (f);
  __m128i i = _mm_cvtps_ph (s, 0);
  _mm_storel_epi64 ((__m128i*)h, i);
}

void
half_to_float4_f16c (const uint16_t h[4],
                     float         f[4])
{
  __m128i i = _mm_loadl_epi64 (CAST_M128I_P (h));
  __m128 s = _mm_cvtph_ps (i);

  _mm_store_ps (f, s);
}

int main(void) {
    float big[4] = { -1.0f, 128.5f, 1.5e4f, 1.5e5f /* too big for fp16 */ };
    uint16_t half[4] = { 0, 0, 0, 0 };

    printf("orig: %f %f %f %f\n", big[0], big[1], big[2], big[3]);

    float_to_half4_f16c(big, half);
    printf("made small...\n");
    printf("half as ints: %x %x %x %x\n", half[0], half[1], half[2], half[3]);

    half_to_float4_f16c(half, big);
    printf("made big again...\n");
    printf("post: %f %f %f %f\n", big[0], big[1], big[2], big[3]);

}

Makefile

.PHONY : clean

borf : borf.c Makefile
	gcc -o borf -mf16c borf.c

clean :
	rm -f borf