/* ------------------------------------------------------------------------- */
int test_sign16s_func(void)
{
INT16 ALIGN(src[65535]), ALIGN(d1[65535]);
#ifdef WITH_SSE2
INT16 ALIGN(d2[65535]);
int i;
#endif
int failed = 0;
char testStr[256];
/* Test when we can reach 16-byte alignment */
testStr[0] = '\0';
get_random_data(src, sizeof(src));
general_sign_16s(src+1, d1+1, 65535);
#ifdef WITH_SSE2
if (IsProcessorFeaturePresentEx(PF_EX_SSSE3))
{
strcat(testStr, " SSSE3");
ssse3_sign_16s(src+1, d2+1, 65535);
for (i=1; i<65535; ++i)
{
if (d1[i] != d2[i])
{
printf("SIGN16s-SSE-aligned FAIL[%d] of %d: want %d, got %d\n",
i, src[i], d1[i], d2[i]);
++failed;
}
}
}
#endif /* i386 */
/* Test when we cannot reach 16-byte alignment */
get_random_data(src, sizeof(src));
general_sign_16s(src+1, d1+2, 65535);
#ifdef WITH_SSE2
if (IsProcessorFeaturePresentEx(PF_EX_SSSE3))
{
ssse3_sign_16s(src+1, d2+2, 65535);
for (i=2; i<65535; ++i)
{
if (d1[i] != d2[i])
{
printf("SIGN16s-SSE-unaligned FAIL[%d] of %d: want %d, got %d\n",
i, src[i-1], d1[i], d2[i]);
++failed;
}
}
}
#endif /* i386 */
if (!failed) printf("All sign16s tests passed (%s).\n", testStr);
return (failed > 0) ? FAILURE : SUCCESS;
}
/* ------------------------------------------------------------------------- */
void primitives_init_sign_opt(primitives_t *prims)
{
/* Pick tuned versions if possible. */
/* I didn't spot an IPP version of this. */
#if defined(WITH_SSE2)
if (IsProcessorFeaturePresentEx(PF_EX_SSSE3)
&& IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
{
prims->sign_16s = ssse3_sign_16s;
}
#endif
}
void primitives_flags_str(char* str, size_t len)
{
int i;
*str = '\0';
--len; /* for the '/0' */
for (i = 0; i < sizeof(flags) / sizeof(flagpair_t); ++i)
{
if (IsProcessorFeaturePresent(flags[i].flag))
{
int slen = strlen(flags[i].str) + 1;
if (len < slen)
break;
if (*str != '\0')
strcat(str, " ");
strcat(str, flags[i].str);
len -= slen;
}
}
for (i = 0; i < sizeof(flags_extended) / sizeof(flagpair_t); ++i)
{
if (IsProcessorFeaturePresentEx(flags_extended[i].flag))
{
int slen = strlen(flags_extended[i].str) + 1;
if (len < slen)
break;
if (*str != '\0')
strcat(str, " ");
strcat(str, flags_extended[i].str);
len -= slen;
}
}
}
int TestCPUFeatures(int argc, char* argv[])
{
printf("Base CPU Flags:\n");
#ifdef _M_IX86_AMD64
printf("\tPF_MMX_INSTRUCTIONS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_MMX_INSTRUCTIONS_AVAILABLE) ? "yes" : "no");
printf("\tPF_XMMI_INSTRUCTIONS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_XMMI_INSTRUCTIONS_AVAILABLE) ? "yes" : "no");
printf("\tPF_XMMI64_INSTRUCTIONS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE) ? "yes" : "no");
printf("\tPF_3DNOW_INSTRUCTIONS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_3DNOW_INSTRUCTIONS_AVAILABLE) ? "yes" : "no");
printf("\tPF_SSE3_INSTRUCTIONS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE) ? "yes" : "no");
printf("\n");
printf("Extended CPU Flags (not found in windows API):\n");
printf("\tPF_EX_3DNOW_PREFETCH: %s\n", IsProcessorFeaturePresentEx(PF_EX_3DNOW_PREFETCH) ? "yes" : "no");
printf("\tPF_EX_SSSE3: %s\n", IsProcessorFeaturePresentEx(PF_EX_SSSE3) ? "yes" : "no");
printf("\tPF_EX_SSE41: %s\n", IsProcessorFeaturePresentEx(PF_EX_SSE41) ? "yes" : "no");
printf("\tPF_EX_SSE42: %s\n", IsProcessorFeaturePresentEx(PF_EX_SSE42) ? "yes" : "no");
printf("\tPF_EX_AVX: %s\n", IsProcessorFeaturePresentEx(PF_EX_AVX) ? "yes" : "no");
printf("\tPF_EX_FMA: %s\n", IsProcessorFeaturePresentEx(PF_EX_FMA) ? "yes" : "no");
printf("\tPF_EX_AVX_AES: %s\n", IsProcessorFeaturePresentEx(PF_EX_AVX_AES) ? "yes" : "no");
#elif defined(_M_ARM)
printf("\tPF_ARM_NEON_INSTRUCTIONS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_ARM_NEON_INSTRUCTIONS_AVAILABLE) ? "yes" : "no");
printf("\tPF_ARM_THUMB: %s\n", IsProcessorFeaturePresent(PF_ARM_THUMB) ? "yes" : "no");
printf("\tPF_ARM_VFP_32_REGISTERS_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_ARM_VFP_32_REGISTERS_AVAILABLE) ? "yes" : "no");
printf("\tPF_ARM_DIVIDE_INSTRUCTION_AVAILABLE: %s\n", IsProcessorFeaturePresent(PF_ARM_DIVIDE_INSTRUCTION_AVAILABLE) ? "yes" : "no");
printf("\tPF_ARM_VFP3: %s\n", IsProcessorFeaturePresent(PF_ARM_VFP3) ? "yes" : "no");
printf("\tPF_ARM_THUMB: %s\n", IsProcessorFeaturePresent(PF_ARM_THUMB) ? "yes" : "no");
printf("\tPF_ARM_JAZELLE: %s\n", IsProcessorFeaturePresent(PF_ARM_JAZELLE) ? "yes" : "no");
printf("\tPF_ARM_DSP: %s\n", IsProcessorFeaturePresent(PF_ARM_DSP) ? "yes" : "no");
printf("\tPF_ARM_THUMB2: %s\n", IsProcessorFeaturePresent(PF_ARM_THUMB2) ? "yes" : "no");
printf("\tPF_ARM_T2EE: %s\n", IsProcessorFeaturePresent(PF_ARM_T2EE) ? "yes" : "no");
printf("\tPF_ARM_INTEL_WMMX: %s\n", IsProcessorFeaturePresent(PF_ARM_INTEL_WMMX) ? "yes" : "no");
printf("Extended CPU Flags (not found in windows API):\n");
printf("\tPF_EX_ARM_VFP1: %s\n", IsProcessorFeaturePresentEx(PF_EX_ARM_VFP1) ? "yes" : "no");
printf("\tPF_EX_ARM_VFP3D16: %s\n", IsProcessorFeaturePresentEx(PF_EX_ARM_VFP3D16) ? "yes" : "no");
printf("\tPF_EX_ARM_VFP4: %s\n", IsProcessorFeaturePresentEx(PF_EX_ARM_VFP4) ? "yes" : "no");
printf("\tPF_EX_ARM_IDIVA: %s\n", IsProcessorFeaturePresentEx(PF_EX_ARM_IDIVA) ? "yes" : "no");
printf("\tPF_EX_ARM_IDIVT: %s\n", IsProcessorFeaturePresentEx(PF_EX_ARM_IDIVT) ? "yes" : "no");
#endif
printf("\n");
return 0;
}
static BOOL CALLBACK rfx_rlgr_init(PINIT_ONCE once, PVOID param, PVOID *context)
{
g_LZCNT = IsProcessorFeaturePresentEx(PF_EX_LZCNT);
return TRUE;
}
