/* ------------------------------------------------------------------------- */
static pstatus_t sse2_set_32s(
INT32 val,
INT32* pDst,
UINT32 len)
{
UINT32 uval = *((UINT32*) &val);
return sse2_set_32u(uval, (UINT32*) pDst, len);
}
/* ------------------------------------------------------------------------- */
int test_set32u_func(void)
{
#if defined(WITH_SSE2) || defined(WITH_IPP)
UINT32 ALIGN(dest[512]);
int off;
#endif
int failed = 0;
char testStr[256];
testStr[0] = '\0';
#ifdef WITH_SSE2
/* Test SSE under various alignments */
if (IsProcessorFeaturePresent(PF_SSE2_INSTRUCTIONS_AVAILABLE))
{
strcat(testStr, " SSE2");
for (off=0; off<16; ++off) {
int len;
for (len=1; len<512-off; ++len)
{
int i;
memset(dest, 0, sizeof(dest));
sse2_set_32u(0xdeadbeefU, dest+off, len);
for (i=0; i<len; ++i)
{
if (dest[off+i] != 0xdeadbeefU)
{
printf("set32u-SSE FAIL: off=%d len=%d dest[%d]=0x%08x\n",
off, len, i+off, dest[i+off]);
failed=1;
}
}
}
}
}
#endif /* i386 */
#ifdef WITH_IPP
strcat(testStr, " IPP");
for (off=0; off<16; ++off) {
int len;
for (len=1; len<512-off; ++len)
{
int i;
memset(dest, 0, sizeof(dest));
ipp_wrapper_set_32u(0xdeadbeefU, dest+off, len);
for (i=0; i<len; ++i)
{
if (dest[off+i] != 0xdeadbeefU)
{
printf("set32u-IPP FAIL: off=%d len=%d dest[%d]=0x%08x\n",
off, len, i+off, dest[i+off]);
failed=1;
}
}
}
}
#endif /* WITH_IPP */
if (!failed) printf("All set32u tests passed (%s).\n", testStr);
return (failed > 0) ? FAILURE : SUCCESS;
}
}
/* ------------------------------------------------------------------------- */
static inline void memset32u_naive(
UINT32 val,
UINT32 *dst,
size_t count)
{
while (count--) *dst++ = val;
}
/* ------------------------------------------------------------------------- */
STD_SPEED_TEST(set32u_speed_test, UINT32, UINT32, dst=dst,
TRUE, memset32u_naive(constant, dst, size),
#ifdef WITH_SSE2
TRUE, sse2_set_32u(constant, dst, size), PF_SSE2_INSTRUCTIONS_AVAILABLE, FALSE,
#else
FALSE, PRIM_NOP, 0, FALSE,
#endif
TRUE, ipp_wrapper_set_32u(constant, dst, size));
int test_set32u_speed(void)
{
UINT32 ALIGN(dst[MAX_TEST_SIZE+1]);
set32u_speed_test("set32u", "aligned", NULL, NULL, 0xdeadbeef, dst,
set_sizes, NUM_SET_SIZES, MEMSET32_PRETEST_ITERATIONS, TEST_TIME);
#if 0
/* Not really necessary; should be almost as fast. */
set32u_speed_test("set32u", "unaligned", NULL, NULL, dst+1,
set_sizes, NUM_SET_SIZES, MEMSET32_PRETEST_ITERATIONS, TEST_TIME);
#endif
