|
Packit |
1fb8d4 |
/* FreeRDP: A Remote Desktop Protocol Client
|
|
Packit |
1fb8d4 |
* Optimized sign operations.
|
|
Packit |
1fb8d4 |
* vi:ts=4 sw=4:
|
|
Packit |
1fb8d4 |
*
|
|
Packit |
1fb8d4 |
* (c) Copyright 2012 Hewlett-Packard Development Company, L.P.
|
|
Packit |
1fb8d4 |
* Licensed under the Apache License, Version 2.0 (the "License"); you may
|
|
Packit |
1fb8d4 |
* not use this file except in compliance with the License. You may obtain
|
|
Packit |
1fb8d4 |
* a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.
|
|
Packit |
1fb8d4 |
* Unless required by applicable law or agreed to in writing, software
|
|
Packit |
1fb8d4 |
* distributed under the License is distributed on an "AS IS" BASIS,
|
|
Packit |
1fb8d4 |
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
|
|
Packit |
1fb8d4 |
* or implied. See the License for the specific language governing
|
|
Packit |
1fb8d4 |
* permissions and limitations under the License.
|
|
Packit |
1fb8d4 |
*/
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
#ifdef HAVE_CONFIG_H
|
|
Packit |
1fb8d4 |
#include "config.h"
|
|
Packit |
1fb8d4 |
#endif
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
#include <freerdp/types.h>
|
|
Packit |
1fb8d4 |
#include <freerdp/primitives.h>
|
|
Packit |
1fb8d4 |
#include <winpr/sysinfo.h>
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
#ifdef WITH_SSE2
|
|
Packit |
1fb8d4 |
#include <emmintrin.h>
|
|
Packit |
1fb8d4 |
#include <tmmintrin.h>
|
|
Packit |
1fb8d4 |
#endif /* WITH_SSE2 */
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
#include "prim_internal.h"
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
static primitives_t* generic = NULL;
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
#ifdef WITH_SSE2
|
|
Packit |
1fb8d4 |
/* ------------------------------------------------------------------------- */
|
|
Packit |
1fb8d4 |
static pstatus_t ssse3_sign_16s(
|
|
Packit |
1fb8d4 |
const INT16* pSrc,
|
|
Packit |
1fb8d4 |
INT16* pDst,
|
|
Packit |
1fb8d4 |
UINT32 len)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
const INT16* sptr = (const INT16*) pSrc;
|
|
Packit |
1fb8d4 |
INT16* dptr = (INT16*) pDst;
|
|
Packit |
1fb8d4 |
size_t count;
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
if (len < 16)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
return generic->sign_16s(pSrc, pDst, len);
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
/* Check for 16-byte alignment (eventually). */
|
|
Packit |
1fb8d4 |
if ((ULONG_PTR) pDst & 0x01)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
return generic->sign_16s(pSrc, pDst, len);
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
/* Seek 16-byte alignment. */
|
|
Packit |
1fb8d4 |
while ((ULONG_PTR) dptr & 0x0f)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
INT16 src = *sptr++;
|
|
Packit |
1fb8d4 |
*dptr++ = (src < 0) ? (-1) : ((src > 0) ? 1 : 0);
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
if (--len == 0) return PRIMITIVES_SUCCESS;
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
/* Do 32-short chunks using 8 XMM registers. */
|
|
Packit |
1fb8d4 |
count = len >> 5; /* / 32 */
|
|
Packit |
1fb8d4 |
len -= count << 5; /* * 32 */
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
if ((ULONG_PTR) sptr & 0x0f)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
/* Unaligned */
|
|
Packit |
1fb8d4 |
while (count--)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
__m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
|
|
Packit |
1fb8d4 |
xmm0 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm1 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm2 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm3 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm4 = _mm_lddqu_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm5 = _mm_lddqu_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm6 = _mm_lddqu_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm7 = _mm_lddqu_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm0 = _mm_sign_epi16(xmm0, xmm4);
|
|
Packit |
1fb8d4 |
xmm1 = _mm_sign_epi16(xmm1, xmm5);
|
|
Packit |
1fb8d4 |
xmm2 = _mm_sign_epi16(xmm2, xmm6);
|
|
Packit |
1fb8d4 |
xmm3 = _mm_sign_epi16(xmm3, xmm7);
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm0);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm1);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm2);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm3);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
else
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
/* Aligned */
|
|
Packit |
1fb8d4 |
while (count--)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
__m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7;
|
|
Packit |
1fb8d4 |
xmm0 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm1 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm2 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm3 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
xmm4 = _mm_load_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm5 = _mm_load_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm6 = _mm_load_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm7 = _mm_load_si128((__m128i*) sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm0 = _mm_sign_epi16(xmm0, xmm4);
|
|
Packit |
1fb8d4 |
xmm1 = _mm_sign_epi16(xmm1, xmm5);
|
|
Packit |
1fb8d4 |
xmm2 = _mm_sign_epi16(xmm2, xmm6);
|
|
Packit |
1fb8d4 |
xmm3 = _mm_sign_epi16(xmm3, xmm7);
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm0);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm1);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm2);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm3);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
/* Do 8-short chunks using two XMM registers. */
|
|
Packit |
1fb8d4 |
count = len >> 3;
|
|
Packit |
1fb8d4 |
len -= count << 3;
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
while (count--)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
__m128i xmm0 = _mm_set1_epi16(0x0001U);
|
|
Packit |
1fb8d4 |
__m128i xmm1 = LOAD_SI128(sptr);
|
|
Packit |
1fb8d4 |
sptr += 8;
|
|
Packit |
1fb8d4 |
xmm0 = _mm_sign_epi16(xmm0, xmm1);
|
|
Packit |
1fb8d4 |
_mm_store_si128((__m128i*) dptr, xmm0);
|
|
Packit |
1fb8d4 |
dptr += 8;
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
/* Do leftovers. */
|
|
Packit |
1fb8d4 |
while (len--)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
INT16 src = *sptr++;
|
|
Packit |
1fb8d4 |
*dptr++ = (src < 0) ? -1 : ((src > 0) ? 1 : 0);
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
return PRIMITIVES_SUCCESS;
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
#endif /* WITH_SSE2 */
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
/* ------------------------------------------------------------------------- */
|
|
Packit |
1fb8d4 |
void primitives_init_sign_opt(primitives_t* prims)
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
generic = primitives_get_generic();
|
|
Packit |
1fb8d4 |
primitives_init_sign(prims);
|
|
Packit |
1fb8d4 |
/* Pick tuned versions if possible. */
|
|
Packit |
1fb8d4 |
/* I didn't spot an IPP version of this. */
|
|
Packit |
1fb8d4 |
#if defined(WITH_SSE2)
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
if (IsProcessorFeaturePresentEx(PF_EX_SSSE3)
|
|
Packit |
1fb8d4 |
&& IsProcessorFeaturePresent(PF_SSE3_INSTRUCTIONS_AVAILABLE))
|
|
Packit |
1fb8d4 |
{
|
|
Packit |
1fb8d4 |
prims->sign_16s = ssse3_sign_16s;
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|
|
Packit |
1fb8d4 |
#endif
|
|
Packit |
1fb8d4 |
}
|
|
Packit |
1fb8d4 |
|