2011-05-25 05:23:39 +02:00
|
|
|
/*
|
|
|
|
|
* software RGB to RGB converter
|
|
|
|
|
* pluralize by software PAL8 to RGB converter
|
|
|
|
|
* software YUV to YUV converter
|
|
|
|
|
* software YUV to RGB converter
|
|
|
|
|
* Written by Nick Kurshev.
|
|
|
|
|
* palette & YUV & runtime CPU stuff by Michael (michaelni@gmx.at)
|
|
|
|
|
* lot of big-endian byte order fixes by Alex Beregszaszi
|
|
|
|
|
*
|
|
|
|
|
* This file is part of FFmpeg.
|
|
|
|
|
*
|
|
|
|
|
* FFmpeg 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.
|
|
|
|
|
*
|
|
|
|
|
* FFmpeg 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 FFmpeg; if not, write to the Free Software
|
|
|
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include <stddef.h>
|
2014-01-09 10:27:20 +01:00
|
|
|
#include <stdint.h>
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2013-04-18 15:54:45 +02:00
|
|
|
#include "libavutil/attributes.h"
|
2014-01-09 10:27:20 +01:00
|
|
|
#include "libavutil/x86/asm.h"
|
2013-04-18 15:54:45 +02:00
|
|
|
|
2011-05-25 05:23:39 +02:00
|
|
|
#undef PREFETCH
|
|
|
|
|
#undef MOVNTQ
|
|
|
|
|
#undef EMMS
|
|
|
|
|
#undef SFENCE
|
|
|
|
|
#undef PAVGB
|
|
|
|
|
|
|
|
|
|
#if COMPILE_TEMPLATE_AMD3DNOW
|
|
|
|
|
#define PREFETCH "prefetch"
|
|
|
|
|
#define PAVGB "pavgusb"
|
2012-07-08 18:42:12 +02:00
|
|
|
#elif COMPILE_TEMPLATE_MMXEXT
|
2011-05-25 05:23:39 +02:00
|
|
|
#define PREFETCH "prefetchnta"
|
|
|
|
|
#define PAVGB "pavgb"
|
|
|
|
|
#else
|
|
|
|
|
#define PREFETCH " # nop"
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
#if COMPILE_TEMPLATE_AMD3DNOW
|
|
|
|
|
/* On K6 femms is faster than emms. On K7 femms is directly mapped to emms. */
|
|
|
|
|
#define EMMS "femms"
|
|
|
|
|
#else
|
|
|
|
|
#define EMMS "emms"
|
|
|
|
|
#endif
|
|
|
|
|
|
2012-07-08 18:42:12 +02:00
|
|
|
#if COMPILE_TEMPLATE_MMXEXT
|
2011-05-25 05:23:39 +02:00
|
|
|
#define MOVNTQ "movntq"
|
|
|
|
|
#define SFENCE "sfence"
|
|
|
|
|
#else
|
|
|
|
|
#define MOVNTQ "movq"
|
|
|
|
|
#define SFENCE " # nop"
|
|
|
|
|
#endif
|
|
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_SSE2
|
|
|
|
|
|
|
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb24tobgr32)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
uint8_t *dest = dst;
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
|
|
|
|
|
mm_end = end - 23;
|
|
|
|
|
__asm__ volatile("movq %0, %%mm7"::"m"(mask32a):"memory");
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 3(%1), %%mm0 \n\t"
|
|
|
|
|
"movd 6(%1), %%mm1 \n\t"
|
|
|
|
|
"punpckldq 9(%1), %%mm1 \n\t"
|
|
|
|
|
"movd 12(%1), %%mm2 \n\t"
|
|
|
|
|
"punpckldq 15(%1), %%mm2 \n\t"
|
|
|
|
|
"movd 18(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 21(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"por %%mm7, %%mm0 \n\t"
|
|
|
|
|
"por %%mm7, %%mm1 \n\t"
|
|
|
|
|
"por %%mm7, %%mm2 \n\t"
|
|
|
|
|
"por %%mm7, %%mm3 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1, 8(%0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 16(%0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 24(%0)"
|
|
|
|
|
:: "r"(dest), "r"(s)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
dest += 32;
|
|
|
|
|
s += 24;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
*dest++ = *s++;
|
|
|
|
|
*dest++ = *s++;
|
|
|
|
|
*dest++ = *s++;
|
|
|
|
|
*dest++ = 255;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
#define STORE_BGR24_MMX \
|
|
|
|
|
"psrlq $8, %%mm2 \n\t" \
|
|
|
|
|
"psrlq $8, %%mm3 \n\t" \
|
|
|
|
|
"psrlq $8, %%mm6 \n\t" \
|
|
|
|
|
"psrlq $8, %%mm7 \n\t" \
|
|
|
|
|
"pand "MANGLE(mask24l)", %%mm0\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24l)", %%mm1\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24l)", %%mm4\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24l)", %%mm5\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24h)", %%mm2\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24h)", %%mm3\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24h)", %%mm6\n\t" \
|
|
|
|
|
"pand "MANGLE(mask24h)", %%mm7\n\t" \
|
|
|
|
|
"por %%mm2, %%mm0 \n\t" \
|
|
|
|
|
"por %%mm3, %%mm1 \n\t" \
|
|
|
|
|
"por %%mm6, %%mm4 \n\t" \
|
|
|
|
|
"por %%mm7, %%mm5 \n\t" \
|
|
|
|
|
\
|
|
|
|
|
"movq %%mm1, %%mm2 \n\t" \
|
|
|
|
|
"movq %%mm4, %%mm3 \n\t" \
|
|
|
|
|
"psllq $48, %%mm2 \n\t" \
|
|
|
|
|
"psllq $32, %%mm3 \n\t" \
|
|
|
|
|
"por %%mm2, %%mm0 \n\t" \
|
|
|
|
|
"psrlq $16, %%mm1 \n\t" \
|
|
|
|
|
"psrlq $32, %%mm4 \n\t" \
|
|
|
|
|
"psllq $16, %%mm5 \n\t" \
|
|
|
|
|
"por %%mm3, %%mm1 \n\t" \
|
|
|
|
|
"por %%mm5, %%mm4 \n\t" \
|
|
|
|
|
\
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t" \
|
|
|
|
|
MOVNTQ" %%mm1, 8(%0) \n\t" \
|
|
|
|
|
MOVNTQ" %%mm4, 16(%0)"
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb32tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
uint8_t *dest = dst;
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
|
|
|
|
|
mm_end = end - 31;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm1 \n\t"
|
|
|
|
|
"movq 16(%1), %%mm4 \n\t"
|
|
|
|
|
"movq 24(%1), %%mm5 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm6 \n\t"
|
|
|
|
|
"movq %%mm5, %%mm7 \n\t"
|
|
|
|
|
STORE_BGR24_MMX
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
:: "r"(dest), "r"(s)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mask24l,mask24h)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
dest += 24;
|
|
|
|
|
s += 32;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
*dest++ = *s++;
|
|
|
|
|
*dest++ = *s++;
|
|
|
|
|
*dest++ = *s++;
|
|
|
|
|
s++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
original by Strepto/Astral
|
|
|
|
|
ported to gcc & bugfixed: A'rpi
|
2012-07-08 19:16:20 +02:00
|
|
|
MMXEXT, 3DNOW optimization by Nick Kurshev
|
2011-05-25 05:23:39 +02:00
|
|
|
32-bit C version, and and&add trick by Michael Niedermayer
|
|
|
|
|
*/
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb15to16)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
register const uint8_t* s=src;
|
|
|
|
|
register uint8_t* d=dst;
|
|
|
|
|
register const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s));
|
|
|
|
|
__asm__ volatile("movq %0, %%mm4"::"m"(mask15s));
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
while (s<mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm4, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm4, %%mm2 \n\t"
|
|
|
|
|
"paddw %%mm1, %%mm0 \n\t"
|
|
|
|
|
"paddw %%mm3, %%mm2 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 8(%0)"
|
|
|
|
|
:: "r"(d), "r"(s)
|
2011-05-25 05:23:39 +02:00
|
|
|
);
|
|
|
|
|
d+=16;
|
|
|
|
|
s+=16;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
mm_end = end - 3;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
register unsigned x= *((const uint32_t *)s);
|
|
|
|
|
*((uint32_t *)d) = (x&0x7FFF7FFF) + (x&0x7FE07FE0);
|
|
|
|
|
d+=4;
|
|
|
|
|
s+=4;
|
|
|
|
|
}
|
|
|
|
|
if (s < end) {
|
|
|
|
|
register unsigned short x= *((const uint16_t *)s);
|
|
|
|
|
*((uint16_t *)d) = (x&0x7FFF) + (x&0x7FE0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb16to15)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
register const uint8_t* s=src;
|
|
|
|
|
register uint8_t* d=dst;
|
|
|
|
|
register const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s));
|
|
|
|
|
__asm__ volatile("movq %0, %%mm7"::"m"(mask15rg));
|
|
|
|
|
__asm__ volatile("movq %0, %%mm6"::"m"(mask15b));
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
while (s<mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlq $1, %%mm0 \n\t"
|
|
|
|
|
"psrlq $1, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm3 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm3, %%mm2 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 8(%0)"
|
|
|
|
|
:: "r"(d), "r"(s)
|
2011-05-25 05:23:39 +02:00
|
|
|
);
|
|
|
|
|
d+=16;
|
|
|
|
|
s+=16;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
mm_end = end - 3;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
register uint32_t x= *((const uint32_t*)s);
|
|
|
|
|
*((uint32_t *)d) = ((x>>1)&0x7FE07FE0) | (x&0x001F001F);
|
|
|
|
|
s+=4;
|
|
|
|
|
d+=4;
|
|
|
|
|
}
|
|
|
|
|
if (s < end) {
|
|
|
|
|
register uint16_t x= *((const uint16_t*)s);
|
|
|
|
|
*((uint16_t *)d) = ((x>>1)&0x7FE0) | (x&0x001F);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb32to16)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %3, %%mm5 \n\t"
|
|
|
|
|
"movq %4, %%mm6 \n\t"
|
|
|
|
|
"movq %5, %%mm7 \n\t"
|
|
|
|
|
"jmp 2f \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 4(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 8(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 12(%1), %%mm3 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm3 \n\t"
|
|
|
|
|
"pmaddwd %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pmaddwd %%mm7, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm5, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm5, %%mm4 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"psrld $5, %%mm0 \n\t"
|
|
|
|
|
"pslld $11, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
"add $16, %1 \n\t"
|
|
|
|
|
"add $8, %0 \n\t"
|
|
|
|
|
"2: \n\t"
|
|
|
|
|
"cmp %2, %1 \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
: "+r" (d), "+r"(s)
|
|
|
|
|
: "r" (mm_end), "m" (mask3216g), "m" (mask3216br), "m" (mul3216)
|
|
|
|
|
);
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register int rgb = *(const uint32_t*)s; s += 4;
|
|
|
|
|
*d++ = ((rgb&0xFF)>>3) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>8);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb32tobgr16)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %0, %%mm7 \n\t"
|
|
|
|
|
"movq %1, %%mm6 \n\t"
|
|
|
|
|
::"m"(red_16mask),"m"(green_16mask));
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 4(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 8(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 12(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"psllq $8, %%mm0 \n\t"
|
|
|
|
|
"psllq $8, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"psrlq $5, %%mm1 \n\t"
|
|
|
|
|
"psrlq $5, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"psrlq $19, %%mm2 \n\t"
|
|
|
|
|
"psrlq $19, %%mm5 \n\t"
|
|
|
|
|
"pand %2, %%mm2 \n\t"
|
|
|
|
|
"pand %2, %%mm5 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $16, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
:: "r"(d),"r"(s),"m"(blue_16mask):"memory");
|
2011-05-25 05:23:39 +02:00
|
|
|
d += 4;
|
|
|
|
|
s += 16;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register int rgb = *(const uint32_t*)s; s += 4;
|
|
|
|
|
*d++ = ((rgb&0xF8)<<8) + ((rgb&0xFC00)>>5) + ((rgb&0xF80000)>>19);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb32to15)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %3, %%mm5 \n\t"
|
|
|
|
|
"movq %4, %%mm6 \n\t"
|
|
|
|
|
"movq %5, %%mm7 \n\t"
|
|
|
|
|
"jmp 2f \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 4(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 8(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 12(%1), %%mm3 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm3 \n\t"
|
|
|
|
|
"pmaddwd %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pmaddwd %%mm7, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm5, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm5, %%mm4 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"psrld $6, %%mm0 \n\t"
|
|
|
|
|
"pslld $10, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
"add $16, %1 \n\t"
|
|
|
|
|
"add $8, %0 \n\t"
|
|
|
|
|
"2: \n\t"
|
|
|
|
|
"cmp %2, %1 \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
: "+r" (d), "+r"(s)
|
|
|
|
|
: "r" (mm_end), "m" (mask3215g), "m" (mask3216br), "m" (mul3215)
|
|
|
|
|
);
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register int rgb = *(const uint32_t*)s; s += 4;
|
|
|
|
|
*d++ = ((rgb&0xFF)>>3) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>9);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb32tobgr15)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %0, %%mm7 \n\t"
|
|
|
|
|
"movq %1, %%mm6 \n\t"
|
|
|
|
|
::"m"(red_15mask),"m"(green_15mask));
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 4(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 8(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 12(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"psllq $7, %%mm0 \n\t"
|
|
|
|
|
"psllq $7, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"psrlq $6, %%mm1 \n\t"
|
|
|
|
|
"psrlq $6, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"psrlq $19, %%mm2 \n\t"
|
|
|
|
|
"psrlq $19, %%mm5 \n\t"
|
|
|
|
|
"pand %2, %%mm2 \n\t"
|
|
|
|
|
"pand %2, %%mm5 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $16, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
::"r"(d),"r"(s),"m"(blue_15mask):"memory");
|
2011-05-25 05:23:39 +02:00
|
|
|
d += 4;
|
|
|
|
|
s += 16;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register int rgb = *(const uint32_t*)s; s += 4;
|
|
|
|
|
*d++ = ((rgb&0xF8)<<7) + ((rgb&0xF800)>>6) + ((rgb&0xF80000)>>19);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb24tobgr16)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %0, %%mm7 \n\t"
|
|
|
|
|
"movq %1, %%mm6 \n\t"
|
|
|
|
|
::"m"(red_16mask),"m"(green_16mask));
|
|
|
|
|
mm_end = end - 11;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 3(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 6(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 9(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"psrlq $3, %%mm0 \n\t"
|
|
|
|
|
"psrlq $3, %%mm3 \n\t"
|
|
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %2, %%mm3 \n\t"
|
|
|
|
|
"psrlq $5, %%mm1 \n\t"
|
|
|
|
|
"psrlq $5, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"psrlq $8, %%mm2 \n\t"
|
|
|
|
|
"psrlq $8, %%mm5 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm5 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $16, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
::"r"(d),"r"(s),"m"(blue_16mask):"memory");
|
2011-05-25 05:23:39 +02:00
|
|
|
d += 4;
|
|
|
|
|
s += 12;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
const int b = *s++;
|
|
|
|
|
const int g = *s++;
|
|
|
|
|
const int r = *s++;
|
|
|
|
|
*d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb24to16)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %0, %%mm7 \n\t"
|
|
|
|
|
"movq %1, %%mm6 \n\t"
|
|
|
|
|
::"m"(red_16mask),"m"(green_16mask));
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 3(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 6(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 9(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"psllq $8, %%mm0 \n\t"
|
|
|
|
|
"psllq $8, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"psrlq $5, %%mm1 \n\t"
|
|
|
|
|
"psrlq $5, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"psrlq $19, %%mm2 \n\t"
|
|
|
|
|
"psrlq $19, %%mm5 \n\t"
|
|
|
|
|
"pand %2, %%mm2 \n\t"
|
|
|
|
|
"pand %2, %%mm5 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $16, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
::"r"(d),"r"(s),"m"(blue_16mask):"memory");
|
2011-05-25 05:23:39 +02:00
|
|
|
d += 4;
|
|
|
|
|
s += 12;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
const int r = *s++;
|
|
|
|
|
const int g = *s++;
|
|
|
|
|
const int b = *s++;
|
|
|
|
|
*d++ = (b>>3) | ((g&0xFC)<<3) | ((r&0xF8)<<8);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb24tobgr15)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %0, %%mm7 \n\t"
|
|
|
|
|
"movq %1, %%mm6 \n\t"
|
|
|
|
|
::"m"(red_15mask),"m"(green_15mask));
|
|
|
|
|
mm_end = end - 11;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 3(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 6(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 9(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"psrlq $3, %%mm0 \n\t"
|
|
|
|
|
"psrlq $3, %%mm3 \n\t"
|
|
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %2, %%mm3 \n\t"
|
|
|
|
|
"psrlq $6, %%mm1 \n\t"
|
|
|
|
|
"psrlq $6, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"psrlq $9, %%mm2 \n\t"
|
|
|
|
|
"psrlq $9, %%mm5 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm5 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $16, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
::"r"(d),"r"(s),"m"(blue_15mask):"memory");
|
2011-05-25 05:23:39 +02:00
|
|
|
d += 4;
|
|
|
|
|
s += 12;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
const int b = *s++;
|
|
|
|
|
const int g = *s++;
|
|
|
|
|
const int r = *s++;
|
|
|
|
|
*d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb24to15)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint8_t *s = src;
|
|
|
|
|
const uint8_t *end;
|
|
|
|
|
const uint8_t *mm_end;
|
|
|
|
|
uint16_t *d = (uint16_t *)dst;
|
|
|
|
|
end = s + src_size;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*src):"memory");
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %0, %%mm7 \n\t"
|
|
|
|
|
"movq %1, %%mm6 \n\t"
|
|
|
|
|
::"m"(red_15mask),"m"(green_15mask));
|
|
|
|
|
mm_end = end - 15;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movd (%1), %%mm0 \n\t"
|
|
|
|
|
"movd 3(%1), %%mm3 \n\t"
|
|
|
|
|
"punpckldq 6(%1), %%mm0 \n\t"
|
|
|
|
|
"punpckldq 9(%1), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"psllq $7, %%mm0 \n\t"
|
|
|
|
|
"psllq $7, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"psrlq $6, %%mm1 \n\t"
|
|
|
|
|
"psrlq $6, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"psrlq $19, %%mm2 \n\t"
|
|
|
|
|
"psrlq $19, %%mm5 \n\t"
|
|
|
|
|
"pand %2, %%mm2 \n\t"
|
|
|
|
|
"pand %2, %%mm5 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $16, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t"
|
|
|
|
|
::"r"(d),"r"(s),"m"(blue_15mask):"memory");
|
2011-05-25 05:23:39 +02:00
|
|
|
d += 4;
|
|
|
|
|
s += 12;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
const int r = *s++;
|
|
|
|
|
const int g = *s++;
|
|
|
|
|
const int b = *s++;
|
|
|
|
|
*d++ = (b>>3) | ((g&0xF8)<<2) | ((r&0xF8)<<7);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb15tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint16_t *end;
|
|
|
|
|
const uint16_t *mm_end;
|
|
|
|
|
uint8_t *d = dst;
|
|
|
|
|
const uint16_t *s = (const uint16_t*)src;
|
|
|
|
|
end = s + src_size/2;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
|
|
|
|
|
mm_end = end - 7;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq (%1), %%mm1 \n\t"
|
|
|
|
|
"movq (%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %3, %%mm1 \n\t"
|
|
|
|
|
"pand %4, %%mm2 \n\t"
|
2011-11-08 16:16:59 +01:00
|
|
|
"psllq $5, %%mm0 \n\t"
|
2012-06-04 21:15:58 +02:00
|
|
|
"pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_mid)", %%mm1 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm5 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm0 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm1 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm2 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm3 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm4 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm5 \n\t"
|
|
|
|
|
"psllq $8, %%mm1 \n\t"
|
|
|
|
|
"psllq $16, %%mm2 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"psllq $8, %%mm4 \n\t"
|
|
|
|
|
"psllq $16, %%mm5 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm0, %%mm6 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm7 \n\t"
|
|
|
|
|
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
"movq 8(%1), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm1 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %3, %%mm1 \n\t"
|
|
|
|
|
"pand %4, %%mm2 \n\t"
|
2011-11-08 16:16:59 +01:00
|
|
|
"psllq $5, %%mm0 \n\t"
|
2012-06-04 21:15:58 +02:00
|
|
|
"pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_mid)", %%mm1 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm5 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm0 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm1 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm2 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm3 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm4 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm5 \n\t"
|
|
|
|
|
"psllq $8, %%mm1 \n\t"
|
|
|
|
|
"psllq $16, %%mm2 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"psllq $8, %%mm4 \n\t"
|
|
|
|
|
"psllq $16, %%mm5 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
|
|
|
|
|
:"=m"(*d)
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
:"r"(s),"m"(mask15b),"m"(mask15g),"m"(mask15r), "m"(mmx_null)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mul15_mid,mul15_hi)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
/* borrowed 32 to 24 */
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %%mm0, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"movq %%mm6, %%mm0 \n\t"
|
|
|
|
|
"movq %%mm7, %%mm1 \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm4, %%mm6 \n\t"
|
|
|
|
|
"movq %%mm5, %%mm7 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t"
|
|
|
|
|
|
|
|
|
|
STORE_BGR24_MMX
|
|
|
|
|
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
:: "r"(d), "m"(*s)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mask24l,mask24h)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
d += 24;
|
|
|
|
|
s += 8;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register uint16_t bgr;
|
|
|
|
|
bgr = *s++;
|
2011-11-08 16:16:59 +01:00
|
|
|
*d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
|
|
|
|
|
*d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
|
|
|
|
|
*d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
|
2011-05-25 05:23:39 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb16tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint16_t *end;
|
|
|
|
|
const uint16_t *mm_end;
|
|
|
|
|
uint8_t *d = (uint8_t *)dst;
|
|
|
|
|
const uint16_t *s = (const uint16_t *)src;
|
|
|
|
|
end = s + src_size/2;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
|
|
|
|
|
mm_end = end - 7;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq (%1), %%mm1 \n\t"
|
|
|
|
|
"movq (%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %3, %%mm1 \n\t"
|
|
|
|
|
"pand %4, %%mm2 \n\t"
|
2011-11-08 16:16:59 +01:00
|
|
|
"psllq $5, %%mm0 \n\t"
|
|
|
|
|
"psrlq $1, %%mm2 \n\t"
|
2012-06-04 21:15:58 +02:00
|
|
|
"pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul16_mid)", %%mm1 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm5 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm0 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm1 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm2 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm3 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm4 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm5 \n\t"
|
|
|
|
|
"psllq $8, %%mm1 \n\t"
|
|
|
|
|
"psllq $16, %%mm2 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"psllq $8, %%mm4 \n\t"
|
|
|
|
|
"psllq $16, %%mm5 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm0, %%mm6 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm7 \n\t"
|
|
|
|
|
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
"movq 8(%1), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm1 \n\t"
|
|
|
|
|
"movq 8(%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %3, %%mm1 \n\t"
|
|
|
|
|
"pand %4, %%mm2 \n\t"
|
2011-11-08 16:16:59 +01:00
|
|
|
"psllq $5, %%mm0 \n\t"
|
|
|
|
|
"psrlq $1, %%mm2 \n\t"
|
2012-06-04 21:15:58 +02:00
|
|
|
"pmulhw "MANGLE(mul15_mid)", %%mm0 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul16_mid)", %%mm1 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm5 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm0 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm1 \n\t"
|
|
|
|
|
"punpcklwd %5, %%mm2 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm3 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm4 \n\t"
|
|
|
|
|
"punpckhwd %5, %%mm5 \n\t"
|
|
|
|
|
"psllq $8, %%mm1 \n\t"
|
|
|
|
|
"psllq $16, %%mm2 \n\t"
|
|
|
|
|
"por %%mm1, %%mm0 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"psllq $8, %%mm4 \n\t"
|
|
|
|
|
"psllq $16, %%mm5 \n\t"
|
|
|
|
|
"por %%mm4, %%mm3 \n\t"
|
|
|
|
|
"por %%mm5, %%mm3 \n\t"
|
|
|
|
|
:"=m"(*d)
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
:"r"(s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mmx_null)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mul15_mid,mul16_mid,mul15_hi)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
/* borrowed 32 to 24 */
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"movq %%mm0, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm3, %%mm5 \n\t"
|
|
|
|
|
"movq %%mm6, %%mm0 \n\t"
|
|
|
|
|
"movq %%mm7, %%mm1 \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm4, %%mm6 \n\t"
|
|
|
|
|
"movq %%mm5, %%mm7 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t"
|
|
|
|
|
|
|
|
|
|
STORE_BGR24_MMX
|
|
|
|
|
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
:: "r"(d), "m"(*s)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mask24l,mask24h)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
d += 24;
|
|
|
|
|
s += 8;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register uint16_t bgr;
|
|
|
|
|
bgr = *s++;
|
2011-11-08 16:16:59 +01:00
|
|
|
*d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
|
|
|
|
|
*d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
|
|
|
|
|
*d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
|
2011-05-25 05:23:39 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* mm0 = 00 B3 00 B2 00 B1 00 B0
|
|
|
|
|
* mm1 = 00 G3 00 G2 00 G1 00 G0
|
|
|
|
|
* mm2 = 00 R3 00 R2 00 R1 00 R0
|
|
|
|
|
* mm6 = FF FF FF FF FF FF FF FF
|
|
|
|
|
* mm7 = 00 00 00 00 00 00 00 00
|
|
|
|
|
*/
|
|
|
|
|
#define PACK_RGB32 \
|
|
|
|
|
"packuswb %%mm7, %%mm0 \n\t" /* 00 00 00 00 B3 B2 B1 B0 */ \
|
|
|
|
|
"packuswb %%mm7, %%mm1 \n\t" /* 00 00 00 00 G3 G2 G1 G0 */ \
|
|
|
|
|
"packuswb %%mm7, %%mm2 \n\t" /* 00 00 00 00 R3 R2 R1 R0 */ \
|
|
|
|
|
"punpcklbw %%mm1, %%mm0 \n\t" /* G3 B3 G2 B2 G1 B1 G0 B0 */ \
|
|
|
|
|
"punpcklbw %%mm6, %%mm2 \n\t" /* FF R3 FF R2 FF R1 FF R0 */ \
|
|
|
|
|
"movq %%mm0, %%mm3 \n\t" \
|
|
|
|
|
"punpcklwd %%mm2, %%mm0 \n\t" /* FF R1 G1 B1 FF R0 G0 B0 */ \
|
|
|
|
|
"punpckhwd %%mm2, %%mm3 \n\t" /* FF R3 G3 B3 FF R2 G2 B2 */ \
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0) \n\t" \
|
|
|
|
|
MOVNTQ" %%mm3, 8(%0) \n\t" \
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb15to32)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint16_t *end;
|
|
|
|
|
const uint16_t *mm_end;
|
|
|
|
|
uint8_t *d = dst;
|
|
|
|
|
const uint16_t *s = (const uint16_t *)src;
|
|
|
|
|
end = s + src_size/2;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
|
|
|
|
|
__asm__ volatile("pxor %%mm7,%%mm7 \n\t":::"memory");
|
|
|
|
|
__asm__ volatile("pcmpeqd %%mm6,%%mm6 \n\t":::"memory");
|
|
|
|
|
mm_end = end - 3;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq (%1), %%mm1 \n\t"
|
|
|
|
|
"movq (%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %3, %%mm1 \n\t"
|
|
|
|
|
"pand %4, %%mm2 \n\t"
|
2011-11-08 16:16:59 +01:00
|
|
|
"psllq $5, %%mm0 \n\t"
|
|
|
|
|
"pmulhw %5, %%mm0 \n\t"
|
|
|
|
|
"pmulhw %5, %%mm1 \n\t"
|
2012-06-04 21:15:58 +02:00
|
|
|
"pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
PACK_RGB32
|
2012-08-14 15:31:43 +02:00
|
|
|
::"r"(d),"r"(s),"m"(mask15b),"m"(mask15g),"m"(mask15r) ,"m"(mul15_mid)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mul15_hi)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
d += 16;
|
|
|
|
|
s += 4;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register uint16_t bgr;
|
|
|
|
|
bgr = *s++;
|
2011-11-08 16:16:59 +01:00
|
|
|
*d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
|
|
|
|
|
*d++ = ((bgr&0x3E0)>>2) | ((bgr&0x3E0)>>7);
|
|
|
|
|
*d++ = ((bgr&0x7C00)>>7) | ((bgr&0x7C00)>>12);
|
2011-05-25 05:23:39 +02:00
|
|
|
*d++ = 255;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb16to32)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
const uint16_t *end;
|
|
|
|
|
const uint16_t *mm_end;
|
|
|
|
|
uint8_t *d = dst;
|
|
|
|
|
const uint16_t *s = (const uint16_t*)src;
|
|
|
|
|
end = s + src_size/2;
|
|
|
|
|
__asm__ volatile(PREFETCH" %0"::"m"(*s):"memory");
|
|
|
|
|
__asm__ volatile("pxor %%mm7,%%mm7 \n\t":::"memory");
|
|
|
|
|
__asm__ volatile("pcmpeqd %%mm6,%%mm6 \n\t":::"memory");
|
|
|
|
|
mm_end = end - 3;
|
|
|
|
|
while (s < mm_end) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1) \n\t"
|
|
|
|
|
"movq (%1), %%mm0 \n\t"
|
|
|
|
|
"movq (%1), %%mm1 \n\t"
|
|
|
|
|
"movq (%1), %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pand %2, %%mm0 \n\t"
|
|
|
|
|
"pand %3, %%mm1 \n\t"
|
|
|
|
|
"pand %4, %%mm2 \n\t"
|
2011-11-08 16:16:59 +01:00
|
|
|
"psllq $5, %%mm0 \n\t"
|
|
|
|
|
"psrlq $1, %%mm2 \n\t"
|
|
|
|
|
"pmulhw %5, %%mm0 \n\t"
|
2012-06-04 21:15:58 +02:00
|
|
|
"pmulhw "MANGLE(mul16_mid)", %%mm1 \n\t"
|
|
|
|
|
"pmulhw "MANGLE(mul15_hi)", %%mm2 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
PACK_RGB32
|
2012-08-14 15:31:43 +02:00
|
|
|
::"r"(d),"r"(s),"m"(mask16b),"m"(mask16g),"m"(mask16r),"m"(mul15_mid)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mul16_mid,mul15_hi)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
d += 16;
|
|
|
|
|
s += 4;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
while (s < end) {
|
|
|
|
|
register uint16_t bgr;
|
|
|
|
|
bgr = *s++;
|
2011-11-08 16:16:59 +01:00
|
|
|
*d++ = ((bgr&0x1F)<<3) | ((bgr&0x1F)>>2);
|
|
|
|
|
*d++ = ((bgr&0x7E0)>>3) | ((bgr&0x7E0)>>9);
|
|
|
|
|
*d++ = ((bgr&0xF800)>>8) | ((bgr&0xF800)>>13);
|
2011-05-25 05:23:39 +02:00
|
|
|
*d++ = 255;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(shuffle_bytes_2103)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
x86_reg idx = 15 - src_size;
|
|
|
|
|
const uint8_t *s = src-idx;
|
|
|
|
|
uint8_t *d = dst-idx;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"test %0, %0 \n\t"
|
|
|
|
|
"jns 2f \n\t"
|
|
|
|
|
PREFETCH" (%1, %0) \n\t"
|
|
|
|
|
"movq %3, %%mm7 \n\t"
|
|
|
|
|
"pxor %4, %%mm7 \n\t"
|
|
|
|
|
"movq %%mm7, %%mm6 \n\t"
|
|
|
|
|
"pxor %5, %%mm7 \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 32(%1, %0) \n\t"
|
|
|
|
|
"movq (%1, %0), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1, %0), %%mm1 \n\t"
|
2012-07-08 18:42:12 +02:00
|
|
|
# if COMPILE_TEMPLATE_MMXEXT
|
2011-05-25 05:23:39 +02:00
|
|
|
"pshufw $177, %%mm0, %%mm3 \n\t"
|
|
|
|
|
"pshufw $177, %%mm1, %%mm5 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm5 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm1 \n\t"
|
|
|
|
|
# else
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm1, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm4 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm5 \n\t"
|
|
|
|
|
"pslld $16, %%mm2 \n\t"
|
|
|
|
|
"psrld $16, %%mm3 \n\t"
|
|
|
|
|
"pslld $16, %%mm4 \n\t"
|
|
|
|
|
"psrld $16, %%mm5 \n\t"
|
|
|
|
|
"por %%mm2, %%mm0 \n\t"
|
|
|
|
|
"por %%mm4, %%mm1 \n\t"
|
|
|
|
|
"por %%mm3, %%mm0 \n\t"
|
|
|
|
|
"por %%mm5, %%mm1 \n\t"
|
|
|
|
|
# endif
|
|
|
|
|
MOVNTQ" %%mm0, (%2, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1, 8(%2, %0) \n\t"
|
|
|
|
|
"add $16, %0 \n\t"
|
|
|
|
|
"js 1b \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
"2: \n\t"
|
|
|
|
|
: "+&r"(idx)
|
|
|
|
|
: "r" (s), "r" (d), "m" (mask32b), "m" (mask32r), "m" (mmx_one)
|
|
|
|
|
: "memory");
|
|
|
|
|
for (; idx<15; idx+=4) {
|
2015-06-23 06:15:07 +02:00
|
|
|
register unsigned v = *(const uint32_t *)&s[idx], g = v & 0xff00ff00;
|
2011-05-25 05:23:39 +02:00
|
|
|
v &= 0xff00ff;
|
|
|
|
|
*(uint32_t *)&d[idx] = (v>>16) + g + (v<<16);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(rgb24tobgr24)(const uint8_t *src, uint8_t *dst, int src_size)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
unsigned i;
|
|
|
|
|
x86_reg mmx_size= 23 - src_size;
|
|
|
|
|
__asm__ volatile (
|
|
|
|
|
"test %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
"jns 2f \n\t"
|
|
|
|
|
"movq "MANGLE(mask24r)", %%mm5 \n\t"
|
|
|
|
|
"movq "MANGLE(mask24g)", %%mm6 \n\t"
|
|
|
|
|
"movq "MANGLE(mask24b)", %%mm7 \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 32(%1, %%"REG_a") \n\t"
|
|
|
|
|
"movq (%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
|
|
|
|
|
"movq (%1, %%"REG_a"), %%mm1 \n\t" // BGR BGR BG
|
|
|
|
|
"movq 2(%1, %%"REG_a"), %%mm2 \n\t" // R BGR BGR B
|
|
|
|
|
"psllq $16, %%mm0 \n\t" // 00 BGR BGR
|
|
|
|
|
"pand %%mm5, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"por %%mm0, %%mm1 \n\t"
|
|
|
|
|
"por %%mm2, %%mm1 \n\t"
|
|
|
|
|
"movq 6(%1, %%"REG_a"), %%mm0 \n\t" // BGR BGR BG
|
|
|
|
|
MOVNTQ" %%mm1, (%2, %%"REG_a") \n\t" // RGB RGB RG
|
|
|
|
|
"movq 8(%1, %%"REG_a"), %%mm1 \n\t" // R BGR BGR B
|
|
|
|
|
"movq 10(%1, %%"REG_a"), %%mm2 \n\t" // GR BGR BGR
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm5, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm6, %%mm2 \n\t"
|
|
|
|
|
"por %%mm0, %%mm1 \n\t"
|
|
|
|
|
"por %%mm2, %%mm1 \n\t"
|
|
|
|
|
"movq 14(%1, %%"REG_a"), %%mm0 \n\t" // R BGR BGR B
|
|
|
|
|
MOVNTQ" %%mm1, 8(%2, %%"REG_a") \n\t" // B RGB RGB R
|
|
|
|
|
"movq 16(%1, %%"REG_a"), %%mm1 \n\t" // GR BGR BGR
|
|
|
|
|
"movq 18(%1, %%"REG_a"), %%mm2 \n\t" // BGR BGR BG
|
|
|
|
|
"pand %%mm6, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm5, %%mm2 \n\t"
|
|
|
|
|
"por %%mm0, %%mm1 \n\t"
|
|
|
|
|
"por %%mm2, %%mm1 \n\t"
|
|
|
|
|
MOVNTQ" %%mm1, 16(%2, %%"REG_a") \n\t"
|
|
|
|
|
"add $24, %%"REG_a" \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
"2: \n\t"
|
|
|
|
|
: "+a" (mmx_size)
|
|
|
|
|
: "r" (src-mmx_size), "r"(dst-mmx_size)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mask24r,mask24g,mask24b)
|
2011-05-25 05:23:39 +02:00
|
|
|
);
|
|
|
|
|
|
|
|
|
|
__asm__ volatile(SFENCE:::"memory");
|
|
|
|
|
__asm__ volatile(EMMS:::"memory");
|
|
|
|
|
|
|
|
|
|
if (mmx_size==23) return; //finished, was multiple of 8
|
|
|
|
|
|
|
|
|
|
src+= src_size;
|
|
|
|
|
dst+= src_size;
|
|
|
|
|
src_size= 23-mmx_size;
|
|
|
|
|
src-= src_size;
|
|
|
|
|
dst-= src_size;
|
|
|
|
|
for (i=0; i<src_size; i+=3) {
|
|
|
|
|
register uint8_t x;
|
|
|
|
|
x = src[i + 2];
|
|
|
|
|
dst[i + 1] = src[i + 1];
|
|
|
|
|
dst[i + 2] = src[i + 0];
|
|
|
|
|
dst[i + 0] = x;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline void RENAME(yuvPlanartoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int dstStride, int vertLumPerChroma)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2011-05-25 05:23:39 +02:00
|
|
|
const x86_reg chromWidth= width>>1;
|
|
|
|
|
for (y=0; y<height; y++) {
|
|
|
|
|
//FIXME handle 2 lines at once (fewer prefetches, reuse some chroma, but very likely memory-limited anyway)
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
PREFETCH" 32(%2, %%"REG_a") \n\t"
|
|
|
|
|
PREFETCH" 32(%3, %%"REG_a") \n\t"
|
|
|
|
|
"movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t" // U(0)
|
|
|
|
|
"movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
|
|
|
|
|
"punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
|
|
|
|
|
"punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
|
|
|
|
|
|
|
|
|
|
"movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
|
|
|
|
|
"movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
|
|
|
|
|
"movq %%mm3, %%mm4 \n\t" // Y(0)
|
|
|
|
|
"movq %%mm5, %%mm6 \n\t" // Y(8)
|
|
|
|
|
"punpcklbw %%mm0, %%mm3 \n\t" // YUYV YUYV(0)
|
|
|
|
|
"punpckhbw %%mm0, %%mm4 \n\t" // YUYV YUYV(4)
|
|
|
|
|
"punpcklbw %%mm2, %%mm5 \n\t" // YUYV YUYV(8)
|
|
|
|
|
"punpckhbw %%mm2, %%mm6 \n\t" // YUYV YUYV(12)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm3, (%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
MOVNTQ" %%mm5, 16(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %4, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
|
|
|
|
|
: "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
|
|
|
|
|
usrc += chromStride;
|
|
|
|
|
vsrc += chromStride;
|
|
|
|
|
}
|
|
|
|
|
ysrc += lumStride;
|
|
|
|
|
dst += dstStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
:::"memory");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Height should be a multiple of 2 and width should be a multiple of 16.
|
|
|
|
|
* (If this is a problem for anyone then tell me, and I will fix it.)
|
|
|
|
|
*/
|
|
|
|
|
static inline void RENAME(yv12toyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
//FIXME interpolate chroma
|
|
|
|
|
RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline void RENAME(yuvPlanartouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int dstStride, int vertLumPerChroma)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2011-05-25 05:23:39 +02:00
|
|
|
const x86_reg chromWidth= width>>1;
|
|
|
|
|
for (y=0; y<height; y++) {
|
|
|
|
|
//FIXME handle 2 lines at once (fewer prefetches, reuse some chroma, but very likely memory-limited anyway)
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 32(%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
PREFETCH" 32(%2, %%"REG_a") \n\t"
|
|
|
|
|
PREFETCH" 32(%3, %%"REG_a") \n\t"
|
|
|
|
|
"movq (%2, %%"REG_a"), %%mm0 \n\t" // U(0)
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t" // U(0)
|
|
|
|
|
"movq (%3, %%"REG_a"), %%mm1 \n\t" // V(0)
|
|
|
|
|
"punpcklbw %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
|
|
|
|
|
"punpckhbw %%mm1, %%mm2 \n\t" // UVUV UVUV(8)
|
|
|
|
|
|
|
|
|
|
"movq (%1, %%"REG_a",2), %%mm3 \n\t" // Y(0)
|
|
|
|
|
"movq 8(%1, %%"REG_a",2), %%mm5 \n\t" // Y(8)
|
|
|
|
|
"movq %%mm0, %%mm4 \n\t" // Y(0)
|
|
|
|
|
"movq %%mm2, %%mm6 \n\t" // Y(8)
|
|
|
|
|
"punpcklbw %%mm3, %%mm0 \n\t" // YUYV YUYV(0)
|
|
|
|
|
"punpckhbw %%mm3, %%mm4 \n\t" // YUYV YUYV(4)
|
|
|
|
|
"punpcklbw %%mm5, %%mm2 \n\t" // YUYV YUYV(8)
|
|
|
|
|
"punpckhbw %%mm5, %%mm6 \n\t" // YUYV YUYV(12)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm0, (%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
MOVNTQ" %%mm4, 8(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 16(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
MOVNTQ" %%mm6, 24(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %4, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
::"r"(dst), "r"(ysrc), "r"(usrc), "r"(vsrc), "g" (chromWidth)
|
|
|
|
|
: "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
|
|
|
|
|
usrc += chromStride;
|
|
|
|
|
vsrc += chromStride;
|
|
|
|
|
}
|
|
|
|
|
ysrc += lumStride;
|
|
|
|
|
dst += dstStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
:::"memory");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Height should be a multiple of 2 and width should be a multiple of 16
|
|
|
|
|
* (If this is a problem for anyone then tell me, and I will fix it.)
|
|
|
|
|
*/
|
|
|
|
|
static inline void RENAME(yv12touyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
//FIXME interpolate chroma
|
|
|
|
|
RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 2);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Width should be a multiple of 16.
|
|
|
|
|
*/
|
|
|
|
|
static inline void RENAME(yuv422ptouyvy)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
RENAME(yuvPlanartouyvy)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Width should be a multiple of 16.
|
|
|
|
|
*/
|
|
|
|
|
static inline void RENAME(yuv422ptoyuy2)(const uint8_t *ysrc, const uint8_t *usrc, const uint8_t *vsrc, uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
RENAME(yuvPlanartoyuy2)(ysrc, usrc, vsrc, dst, width, height, lumStride, chromStride, dstStride, 1);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Height should be a multiple of 2 and width should be a multiple of 16.
|
|
|
|
|
* (If this is a problem for anyone then tell me, and I will fix it.)
|
|
|
|
|
*/
|
|
|
|
|
static inline void RENAME(yuy2toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int srcStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2011-05-25 05:23:39 +02:00
|
|
|
const x86_reg chromWidth= width>>1;
|
|
|
|
|
for (y=0; y<height; y+=2) {
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
"movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
|
|
|
|
|
"movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t" // YUYV YUYV(0)
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t" // YUYV YUYV(4)
|
|
|
|
|
"psrlw $8, %%mm0 \n\t" // U0V0 U0V0(0)
|
|
|
|
|
"psrlw $8, %%mm1 \n\t" // U0V0 U0V0(4)
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm2, (%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
|
|
|
|
|
"movq 16(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(8)
|
|
|
|
|
"movq 24(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(12)
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t" // YUYV YUYV(8)
|
|
|
|
|
"movq %%mm2, %%mm4 \n\t" // YUYV YUYV(12)
|
|
|
|
|
"psrlw $8, %%mm1 \n\t" // U0V0 U0V0(8)
|
|
|
|
|
"psrlw $8, %%mm2 \n\t" // U0V0 U0V0(12)
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
|
|
|
|
|
"pand %%mm7, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
|
|
|
|
|
"packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
|
|
|
|
|
"packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
|
|
|
|
|
"psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
|
|
|
|
|
"psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
|
|
|
|
|
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %4, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
|
|
|
|
|
: "memory", "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
|
|
|
|
|
ydst += lumStride;
|
|
|
|
|
src += srcStride;
|
|
|
|
|
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
"movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
|
|
|
|
|
"movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
|
|
|
|
|
"movq 16(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(8)
|
|
|
|
|
"movq 24(%0, %%"REG_a", 4), %%mm3 \n\t" // YUYV YUYV(12)
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm0, (%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %4, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
|
|
|
|
|
::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
|
|
|
|
|
: "memory", "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
udst += chromStride;
|
|
|
|
|
vdst += chromStride;
|
|
|
|
|
ydst += lumStride;
|
|
|
|
|
src += srcStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
:::"memory");
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2012-07-08 18:42:12 +02:00
|
|
|
#if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-26 18:57:26 +02:00
|
|
|
static inline void RENAME(planar2x)(const uint8_t *src, uint8_t *dst, int srcWidth, int srcHeight, int srcStride, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int x,y;
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
dst[0]= src[0];
|
|
|
|
|
|
|
|
|
|
// first line
|
|
|
|
|
for (x=0; x<srcWidth-1; x++) {
|
|
|
|
|
dst[2*x+1]= (3*src[x] + src[x+1])>>2;
|
|
|
|
|
dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
|
|
|
|
|
}
|
|
|
|
|
dst[2*srcWidth-1]= src[srcWidth-1];
|
|
|
|
|
|
|
|
|
|
dst+= dstStride;
|
|
|
|
|
|
|
|
|
|
for (y=1; y<srcHeight; y++) {
|
|
|
|
|
const x86_reg mmxSize= srcWidth&~15;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"mov %4, %%"REG_a" \n\t"
|
|
|
|
|
"movq "MANGLE(mmx_ff)", %%mm0 \n\t"
|
|
|
|
|
"movq (%0, %%"REG_a"), %%mm4 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm2 \n\t"
|
|
|
|
|
"psllq $8, %%mm4 \n\t"
|
|
|
|
|
"pand %%mm0, %%mm2 \n\t"
|
|
|
|
|
"por %%mm2, %%mm4 \n\t"
|
|
|
|
|
"movq (%1, %%"REG_a"), %%mm5 \n\t"
|
|
|
|
|
"movq %%mm5, %%mm3 \n\t"
|
|
|
|
|
"psllq $8, %%mm5 \n\t"
|
|
|
|
|
"pand %%mm0, %%mm3 \n\t"
|
|
|
|
|
"por %%mm3, %%mm5 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq (%0, %%"REG_a"), %%mm0 \n\t"
|
|
|
|
|
"movq (%1, %%"REG_a"), %%mm1 \n\t"
|
|
|
|
|
"movq 1(%0, %%"REG_a"), %%mm2 \n\t"
|
|
|
|
|
"movq 1(%1, %%"REG_a"), %%mm3 \n\t"
|
|
|
|
|
PAVGB" %%mm0, %%mm5 \n\t"
|
|
|
|
|
PAVGB" %%mm0, %%mm3 \n\t"
|
|
|
|
|
PAVGB" %%mm0, %%mm5 \n\t"
|
|
|
|
|
PAVGB" %%mm0, %%mm3 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm4 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm2 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm4 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm5, %%mm7 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm6 \n\t"
|
|
|
|
|
"punpcklbw %%mm3, %%mm5 \n\t"
|
|
|
|
|
"punpckhbw %%mm3, %%mm7 \n\t"
|
|
|
|
|
"punpcklbw %%mm2, %%mm4 \n\t"
|
|
|
|
|
"punpckhbw %%mm2, %%mm6 \n\t"
|
|
|
|
|
MOVNTQ" %%mm5, (%2, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm7, 8(%2, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm4, (%3, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm6, 8(%3, %%"REG_a", 2) \n\t"
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"movq -1(%0, %%"REG_a"), %%mm4 \n\t"
|
|
|
|
|
"movq -1(%1, %%"REG_a"), %%mm5 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
:: "r" (src + mmxSize ), "r" (src + srcStride + mmxSize ),
|
|
|
|
|
"r" (dst + mmxSize*2), "r" (dst + dstStride + mmxSize*2),
|
|
|
|
|
"g" (-mmxSize)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(mmx_ff)
|
2011-05-25 05:23:39 +02:00
|
|
|
: "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
|
|
|
|
|
for (x=mmxSize-1; x<srcWidth-1; x++) {
|
|
|
|
|
dst[2*x +1]= (3*src[x+0] + src[x+srcStride+1])>>2;
|
|
|
|
|
dst[2*x+dstStride+2]= ( src[x+0] + 3*src[x+srcStride+1])>>2;
|
|
|
|
|
dst[2*x+dstStride+1]= ( src[x+1] + 3*src[x+srcStride ])>>2;
|
|
|
|
|
dst[2*x +2]= (3*src[x+1] + src[x+srcStride ])>>2;
|
|
|
|
|
}
|
|
|
|
|
dst[srcWidth*2 -1 ]= (3*src[srcWidth-1] + src[srcWidth-1 + srcStride])>>2;
|
|
|
|
|
dst[srcWidth*2 -1 + dstStride]= ( src[srcWidth-1] + 3*src[srcWidth-1 + srcStride])>>2;
|
|
|
|
|
|
|
|
|
|
dst+=dstStride*2;
|
|
|
|
|
src+=srcStride;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// last line
|
|
|
|
|
dst[0]= src[0];
|
|
|
|
|
|
|
|
|
|
for (x=0; x<srcWidth-1; x++) {
|
|
|
|
|
dst[2*x+1]= (3*src[x] + src[x+1])>>2;
|
|
|
|
|
dst[2*x+2]= ( src[x] + 3*src[x+1])>>2;
|
|
|
|
|
}
|
|
|
|
|
dst[2*srcWidth-1]= src[srcWidth-1];
|
|
|
|
|
|
|
|
|
|
__asm__ volatile(EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
:::"memory");
|
|
|
|
|
}
|
2012-07-08 18:42:12 +02:00
|
|
|
#endif /* COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
/**
|
|
|
|
|
* Height should be a multiple of 2 and width should be a multiple of 16.
|
|
|
|
|
* (If this is a problem for anyone then tell me, and I will fix it.)
|
|
|
|
|
* Chrominance data is only taken from every second line, others are ignored.
|
|
|
|
|
* FIXME: Write HQ version.
|
|
|
|
|
*/
|
|
|
|
|
static inline void RENAME(uyvytoyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int srcStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2011-05-25 05:23:39 +02:00
|
|
|
const x86_reg chromWidth= width>>1;
|
|
|
|
|
for (y=0; y<height; y+=2) {
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t" // FF,00,FF,00...
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
"movq (%0, %%"REG_a", 4), %%mm0 \n\t" // UYVY UYVY(0)
|
|
|
|
|
"movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // UYVY UYVY(4)
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t" // UYVY UYVY(0)
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t" // UYVY UYVY(4)
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t" // U0V0 U0V0(0)
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(4)
|
|
|
|
|
"psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(0)
|
|
|
|
|
"psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(4)
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t" // UVUV UVUV(0)
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(0)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm2, (%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
|
|
|
|
|
"movq 16(%0, %%"REG_a", 4), %%mm1 \n\t" // UYVY UYVY(8)
|
|
|
|
|
"movq 24(%0, %%"REG_a", 4), %%mm2 \n\t" // UYVY UYVY(12)
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t" // UYVY UYVY(8)
|
|
|
|
|
"movq %%mm2, %%mm4 \n\t" // UYVY UYVY(12)
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t" // U0V0 U0V0(8)
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t" // U0V0 U0V0(12)
|
|
|
|
|
"psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(8)
|
|
|
|
|
"psrlw $8, %%mm4 \n\t" // Y0Y0 Y0Y0(12)
|
|
|
|
|
"packuswb %%mm2, %%mm1 \n\t" // UVUV UVUV(8)
|
|
|
|
|
"packuswb %%mm4, %%mm3 \n\t" // YYYY YYYY(8)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm3, 8(%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm0, %%mm2 \n\t" // UVUV UVUV(0)
|
|
|
|
|
"movq %%mm1, %%mm3 \n\t" // UVUV UVUV(8)
|
|
|
|
|
"psrlw $8, %%mm0 \n\t" // V0V0 V0V0(0)
|
|
|
|
|
"psrlw $8, %%mm1 \n\t" // V0V0 V0V0(8)
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t" // U0U0 U0U0(0)
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t" // U0U0 U0U0(8)
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t" // VVVV VVVV(0)
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t" // UUUU UUUU(0)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm0, (%3, %%"REG_a") \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, (%2, %%"REG_a") \n\t"
|
|
|
|
|
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %4, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
|
|
|
|
|
: "memory", "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
|
|
|
|
|
ydst += lumStride;
|
|
|
|
|
src += srcStride;
|
|
|
|
|
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%0, %%"REG_a", 4) \n\t"
|
|
|
|
|
"movq (%0, %%"REG_a", 4), %%mm0 \n\t" // YUYV YUYV(0)
|
|
|
|
|
"movq 8(%0, %%"REG_a", 4), %%mm1 \n\t" // YUYV YUYV(4)
|
|
|
|
|
"movq 16(%0, %%"REG_a", 4), %%mm2 \n\t" // YUYV YUYV(8)
|
|
|
|
|
"movq 24(%0, %%"REG_a", 4), %%mm3 \n\t" // YUYV YUYV(12)
|
|
|
|
|
"psrlw $8, %%mm0 \n\t" // Y0Y0 Y0Y0(0)
|
|
|
|
|
"psrlw $8, %%mm1 \n\t" // Y0Y0 Y0Y0(4)
|
|
|
|
|
"psrlw $8, %%mm2 \n\t" // Y0Y0 Y0Y0(8)
|
|
|
|
|
"psrlw $8, %%mm3 \n\t" // Y0Y0 Y0Y0(12)
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t" // YYYY YYYY(0)
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t" // YYYY YYYY(8)
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm0, (%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 8(%1, %%"REG_a", 2) \n\t"
|
|
|
|
|
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %4, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
|
|
|
|
|
::"r"(src), "r"(ydst), "r"(udst), "r"(vdst), "g" (chromWidth)
|
|
|
|
|
: "memory", "%"REG_a
|
|
|
|
|
);
|
|
|
|
|
udst += chromStride;
|
|
|
|
|
vdst += chromStride;
|
|
|
|
|
ydst += lumStride;
|
|
|
|
|
src += srcStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__ volatile(EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
:::"memory");
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Height should be a multiple of 2 and width should be a multiple of 2.
|
|
|
|
|
* (If this is a problem for anyone then tell me, and I will fix it.)
|
|
|
|
|
* Chrominance data is only taken from every second line,
|
|
|
|
|
* others are ignored in the C version.
|
|
|
|
|
* FIXME: Write HQ version.
|
|
|
|
|
*/
|
2013-04-18 12:47:16 +02:00
|
|
|
#if HAVE_7REGS
|
2011-05-25 05:23:39 +02:00
|
|
|
static inline void RENAME(rgb24toyv12)(const uint8_t *src, uint8_t *ydst, uint8_t *udst, uint8_t *vdst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
2013-04-15 03:01:04 +02:00
|
|
|
int lumStride, int chromStride, int srcStride,
|
|
|
|
|
int32_t *rgb2yuv)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2013-04-15 21:26:06 +02:00
|
|
|
#define BGR2Y_IDX "16*4+16*32"
|
|
|
|
|
#define BGR2U_IDX "16*4+16*33"
|
|
|
|
|
#define BGR2V_IDX "16*4+16*34"
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2011-05-25 05:23:39 +02:00
|
|
|
const x86_reg chromWidth= width>>1;
|
2014-11-20 00:43:45 +01:00
|
|
|
|
|
|
|
|
if (height > 2) {
|
|
|
|
|
ff_rgb24toyv12_c(src, ydst, udst, vdst, width, 2, lumStride, chromStride, srcStride, rgb2yuv);
|
|
|
|
|
src += 2*srcStride;
|
|
|
|
|
ydst += 2*lumStride;
|
|
|
|
|
udst += chromStride;
|
|
|
|
|
vdst += chromStride;
|
|
|
|
|
height -= 2;
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-25 05:23:39 +02:00
|
|
|
for (y=0; y<height-2; y+=2) {
|
2011-05-26 18:57:26 +02:00
|
|
|
int i;
|
2011-05-25 05:23:39 +02:00
|
|
|
for (i=0; i<2; i++) {
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"mov %2, %%"REG_a" \n\t"
|
2013-04-15 21:26:06 +02:00
|
|
|
"movq "BGR2Y_IDX"(%3), %%mm6 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq "MANGLE(ff_w1111)", %%mm5 \n\t"
|
|
|
|
|
"pxor %%mm7, %%mm7 \n\t"
|
|
|
|
|
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%0, %%"REG_d") \n\t"
|
|
|
|
|
"movd (%0, %%"REG_d"), %%mm0 \n\t"
|
|
|
|
|
"movd 3(%0, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm0 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm1 \n\t"
|
|
|
|
|
"movd 6(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movd 9(%0, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm3 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm0 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm2 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm3 \n\t"
|
|
|
|
|
"psrad $8, %%mm0 \n\t"
|
|
|
|
|
"psrad $8, %%mm1 \n\t"
|
|
|
|
|
"psrad $8, %%mm2 \n\t"
|
|
|
|
|
"psrad $8, %%mm3 \n\t"
|
|
|
|
|
"packssdw %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packssdw %%mm3, %%mm2 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm0 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm2 \n\t"
|
|
|
|
|
"packssdw %%mm2, %%mm0 \n\t"
|
|
|
|
|
"psraw $7, %%mm0 \n\t"
|
|
|
|
|
|
|
|
|
|
"movd 12(%0, %%"REG_d"), %%mm4 \n\t"
|
|
|
|
|
"movd 15(%0, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm4 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm1 \n\t"
|
|
|
|
|
"movd 18(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movd 21(%0, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm3 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm4 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm1 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm2 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm3 \n\t"
|
|
|
|
|
"psrad $8, %%mm4 \n\t"
|
|
|
|
|
"psrad $8, %%mm1 \n\t"
|
|
|
|
|
"psrad $8, %%mm2 \n\t"
|
|
|
|
|
"psrad $8, %%mm3 \n\t"
|
|
|
|
|
"packssdw %%mm1, %%mm4 \n\t"
|
|
|
|
|
"packssdw %%mm3, %%mm2 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm4 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm2 \n\t"
|
|
|
|
|
"add $24, %%"REG_d" \n\t"
|
|
|
|
|
"packssdw %%mm2, %%mm4 \n\t"
|
|
|
|
|
"psraw $7, %%mm4 \n\t"
|
|
|
|
|
|
|
|
|
|
"packuswb %%mm4, %%mm0 \n\t"
|
|
|
|
|
"paddusb "MANGLE(ff_bgr2YOffset)", %%mm0 \n\t"
|
|
|
|
|
|
|
|
|
|
MOVNTQ" %%mm0, (%1, %%"REG_a") \n\t"
|
|
|
|
|
"add $8, %%"REG_a" \n\t"
|
|
|
|
|
" js 1b \n\t"
|
2013-04-15 21:26:06 +02:00
|
|
|
: : "r" (src+width*3), "r" (ydst+width), "g" ((x86_reg)-width), "r"(rgb2yuv)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(ff_w1111,ff_bgr2YOffset)
|
2011-05-25 05:23:39 +02:00
|
|
|
: "%"REG_a, "%"REG_d
|
|
|
|
|
);
|
|
|
|
|
ydst += lumStride;
|
|
|
|
|
src += srcStride;
|
|
|
|
|
}
|
|
|
|
|
src -= srcStride*2;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"mov %4, %%"REG_a" \n\t"
|
|
|
|
|
"movq "MANGLE(ff_w1111)", %%mm5 \n\t"
|
2013-04-15 21:26:06 +02:00
|
|
|
"movq "BGR2U_IDX"(%5), %%mm6 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"pxor %%mm7, %%mm7 \n\t"
|
|
|
|
|
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t"
|
|
|
|
|
"add %%"REG_d", %%"REG_d" \n\t"
|
|
|
|
|
".p2align 4 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%0, %%"REG_d") \n\t"
|
|
|
|
|
PREFETCH" 64(%1, %%"REG_d") \n\t"
|
2012-07-08 18:42:12 +02:00
|
|
|
#if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq (%0, %%"REG_d"), %%mm0 \n\t"
|
|
|
|
|
"movq (%1, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"movq 6(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movq 6(%1, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm0 \n\t"
|
|
|
|
|
PAVGB" %%mm3, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlq $24, %%mm0 \n\t"
|
|
|
|
|
"psrlq $24, %%mm2 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm0 \n\t"
|
|
|
|
|
PAVGB" %%mm3, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm0 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
#else
|
|
|
|
|
"movd (%0, %%"REG_d"), %%mm0 \n\t"
|
|
|
|
|
"movd (%1, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"movd 3(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movd 3(%1, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm0 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm3 \n\t"
|
|
|
|
|
"paddw %%mm1, %%mm0 \n\t"
|
|
|
|
|
"paddw %%mm3, %%mm2 \n\t"
|
|
|
|
|
"paddw %%mm2, %%mm0 \n\t"
|
|
|
|
|
"movd 6(%0, %%"REG_d"), %%mm4 \n\t"
|
|
|
|
|
"movd 6(%1, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"movd 9(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movd 9(%1, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm4 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm3 \n\t"
|
|
|
|
|
"paddw %%mm1, %%mm4 \n\t"
|
|
|
|
|
"paddw %%mm3, %%mm2 \n\t"
|
|
|
|
|
"paddw %%mm4, %%mm2 \n\t"
|
|
|
|
|
"psrlw $2, %%mm0 \n\t"
|
|
|
|
|
"psrlw $2, %%mm2 \n\t"
|
|
|
|
|
#endif
|
2013-04-15 21:26:06 +02:00
|
|
|
"movq "BGR2V_IDX"(%5), %%mm1 \n\t"
|
|
|
|
|
"movq "BGR2V_IDX"(%5), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
"pmaddwd %%mm0, %%mm1 \n\t"
|
|
|
|
|
"pmaddwd %%mm2, %%mm3 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm0 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm2 \n\t"
|
|
|
|
|
"psrad $8, %%mm0 \n\t"
|
|
|
|
|
"psrad $8, %%mm1 \n\t"
|
|
|
|
|
"psrad $8, %%mm2 \n\t"
|
|
|
|
|
"psrad $8, %%mm3 \n\t"
|
|
|
|
|
"packssdw %%mm2, %%mm0 \n\t"
|
|
|
|
|
"packssdw %%mm3, %%mm1 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm0 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm1 \n\t"
|
|
|
|
|
"packssdw %%mm1, %%mm0 \n\t" // V1 V0 U1 U0
|
|
|
|
|
"psraw $7, %%mm0 \n\t"
|
|
|
|
|
|
2012-07-08 18:42:12 +02:00
|
|
|
#if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq 12(%0, %%"REG_d"), %%mm4 \n\t"
|
|
|
|
|
"movq 12(%1, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"movq 18(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movq 18(%1, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm4 \n\t"
|
|
|
|
|
PAVGB" %%mm3, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlq $24, %%mm4 \n\t"
|
|
|
|
|
"psrlq $24, %%mm2 \n\t"
|
|
|
|
|
PAVGB" %%mm1, %%mm4 \n\t"
|
|
|
|
|
PAVGB" %%mm3, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm4 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
#else
|
|
|
|
|
"movd 12(%0, %%"REG_d"), %%mm4 \n\t"
|
|
|
|
|
"movd 12(%1, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"movd 15(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movd 15(%1, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm4 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm3 \n\t"
|
|
|
|
|
"paddw %%mm1, %%mm4 \n\t"
|
|
|
|
|
"paddw %%mm3, %%mm2 \n\t"
|
|
|
|
|
"paddw %%mm2, %%mm4 \n\t"
|
|
|
|
|
"movd 18(%0, %%"REG_d"), %%mm5 \n\t"
|
|
|
|
|
"movd 18(%1, %%"REG_d"), %%mm1 \n\t"
|
|
|
|
|
"movd 21(%0, %%"REG_d"), %%mm2 \n\t"
|
|
|
|
|
"movd 21(%1, %%"REG_d"), %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm5 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm2 \n\t"
|
|
|
|
|
"punpcklbw %%mm7, %%mm3 \n\t"
|
|
|
|
|
"paddw %%mm1, %%mm5 \n\t"
|
|
|
|
|
"paddw %%mm3, %%mm2 \n\t"
|
|
|
|
|
"paddw %%mm5, %%mm2 \n\t"
|
|
|
|
|
"movq "MANGLE(ff_w1111)", %%mm5 \n\t"
|
|
|
|
|
"psrlw $2, %%mm4 \n\t"
|
|
|
|
|
"psrlw $2, %%mm2 \n\t"
|
|
|
|
|
#endif
|
2013-04-15 21:26:06 +02:00
|
|
|
"movq "BGR2V_IDX"(%5), %%mm1 \n\t"
|
|
|
|
|
"movq "BGR2V_IDX"(%5), %%mm3 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
"pmaddwd %%mm4, %%mm1 \n\t"
|
|
|
|
|
"pmaddwd %%mm2, %%mm3 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm4 \n\t"
|
|
|
|
|
"pmaddwd %%mm6, %%mm2 \n\t"
|
|
|
|
|
"psrad $8, %%mm4 \n\t"
|
|
|
|
|
"psrad $8, %%mm1 \n\t"
|
|
|
|
|
"psrad $8, %%mm2 \n\t"
|
|
|
|
|
"psrad $8, %%mm3 \n\t"
|
|
|
|
|
"packssdw %%mm2, %%mm4 \n\t"
|
|
|
|
|
"packssdw %%mm3, %%mm1 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm4 \n\t"
|
|
|
|
|
"pmaddwd %%mm5, %%mm1 \n\t"
|
|
|
|
|
"add $24, %%"REG_d" \n\t"
|
|
|
|
|
"packssdw %%mm1, %%mm4 \n\t" // V3 V2 U3 U2
|
|
|
|
|
"psraw $7, %%mm4 \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"punpckldq %%mm4, %%mm0 \n\t"
|
|
|
|
|
"punpckhdq %%mm4, %%mm1 \n\t"
|
|
|
|
|
"packsswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"paddb "MANGLE(ff_bgr2UVOffset)", %%mm0 \n\t"
|
|
|
|
|
"movd %%mm0, (%2, %%"REG_a") \n\t"
|
|
|
|
|
"punpckhdq %%mm0, %%mm0 \n\t"
|
|
|
|
|
"movd %%mm0, (%3, %%"REG_a") \n\t"
|
|
|
|
|
"add $4, %%"REG_a" \n\t"
|
|
|
|
|
" js 1b \n\t"
|
2013-04-15 21:26:06 +02:00
|
|
|
: : "r" (src+chromWidth*6), "r" (src+srcStride+chromWidth*6), "r" (udst+chromWidth), "r" (vdst+chromWidth), "g" (-chromWidth), "r"(rgb2yuv)
|
2014-03-18 05:53:26 +01:00
|
|
|
NAMED_CONSTRAINTS_ADD(ff_w1111,ff_bgr2UVOffset)
|
2011-05-25 05:23:39 +02:00
|
|
|
: "%"REG_a, "%"REG_d
|
|
|
|
|
);
|
|
|
|
|
|
|
|
|
|
udst += chromStride;
|
|
|
|
|
vdst += chromStride;
|
|
|
|
|
src += srcStride*2;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
__asm__ volatile(EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
:::"memory");
|
|
|
|
|
|
2013-04-15 03:01:04 +02:00
|
|
|
ff_rgb24toyv12_c(src, ydst, udst, vdst, width, height-y, lumStride, chromStride, srcStride, rgb2yuv);
|
2011-05-25 05:23:39 +02:00
|
|
|
}
|
2013-04-18 12:47:16 +02:00
|
|
|
#endif /* HAVE_7REGS */
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_SSE2 */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2013-11-19 14:46:01 +01:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW && !COMPILE_TEMPLATE_AVX
|
2011-05-25 05:23:39 +02:00
|
|
|
static void RENAME(interleaveBytes)(const uint8_t *src1, const uint8_t *src2, uint8_t *dest,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height, int src1Stride,
|
|
|
|
|
int src2Stride, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int h;
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
for (h=0; h < height; h++) {
|
2011-05-26 18:57:26 +02:00
|
|
|
int w;
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2015-12-15 02:50:20 +01:00
|
|
|
if (width >= 16) {
|
2011-05-25 05:23:39 +02:00
|
|
|
#if COMPILE_TEMPLATE_SSE2
|
2015-12-15 02:50:20 +01:00
|
|
|
if (!((((intptr_t)src1) | ((intptr_t)src2) | ((intptr_t)dest))&15)) {
|
2011-05-25 05:23:39 +02:00
|
|
|
__asm__(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%1, %%"REG_a") \n\t"
|
|
|
|
|
PREFETCH" 64(%2, %%"REG_a") \n\t"
|
|
|
|
|
"movdqa (%1, %%"REG_a"), %%xmm0 \n\t"
|
|
|
|
|
"movdqa (%1, %%"REG_a"), %%xmm1 \n\t"
|
|
|
|
|
"movdqa (%2, %%"REG_a"), %%xmm2 \n\t"
|
|
|
|
|
"punpcklbw %%xmm2, %%xmm0 \n\t"
|
|
|
|
|
"punpckhbw %%xmm2, %%xmm1 \n\t"
|
|
|
|
|
"movntdq %%xmm0, (%0, %%"REG_a", 2) \n\t"
|
|
|
|
|
"movntdq %%xmm1, 16(%0, %%"REG_a", 2) \n\t"
|
|
|
|
|
"add $16, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %3, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
::"r"(dest), "r"(src1), "r"(src2), "r" ((x86_reg)width-15)
|
2015-06-23 06:14:16 +02:00
|
|
|
: "memory", XMM_CLOBBERS("xmm0", "xmm1", "xmm2",) "%"REG_a
|
2011-05-25 05:23:39 +02:00
|
|
|
);
|
2015-12-15 02:50:20 +01:00
|
|
|
} else
|
|
|
|
|
#endif
|
2011-05-25 05:23:39 +02:00
|
|
|
__asm__(
|
|
|
|
|
"xor %%"REG_a", %%"REG_a" \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
PREFETCH" 64(%1, %%"REG_a") \n\t"
|
|
|
|
|
PREFETCH" 64(%2, %%"REG_a") \n\t"
|
|
|
|
|
"movq (%1, %%"REG_a"), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1, %%"REG_a"), %%mm2 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"movq (%2, %%"REG_a"), %%mm4 \n\t"
|
|
|
|
|
"movq 8(%2, %%"REG_a"), %%mm5 \n\t"
|
|
|
|
|
"punpcklbw %%mm4, %%mm0 \n\t"
|
|
|
|
|
"punpckhbw %%mm4, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm5, %%mm2 \n\t"
|
|
|
|
|
"punpckhbw %%mm5, %%mm3 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0, (%0, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1, 8(%0, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 16(%0, %%"REG_a", 2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 24(%0, %%"REG_a", 2) \n\t"
|
|
|
|
|
"add $16, %%"REG_a" \n\t"
|
|
|
|
|
"cmp %3, %%"REG_a" \n\t"
|
|
|
|
|
" jb 1b \n\t"
|
|
|
|
|
::"r"(dest), "r"(src1), "r"(src2), "r" ((x86_reg)width-15)
|
|
|
|
|
: "memory", "%"REG_a
|
|
|
|
|
);
|
2015-12-15 02:50:20 +01:00
|
|
|
|
|
|
|
|
}
|
2011-05-25 05:23:39 +02:00
|
|
|
for (w= (width&(~15)); w < width; w++) {
|
|
|
|
|
dest[2*w+0] = src1[w];
|
|
|
|
|
dest[2*w+1] = src2[w];
|
|
|
|
|
}
|
|
|
|
|
dest += dstStride;
|
|
|
|
|
src1 += src1Stride;
|
|
|
|
|
src2 += src2Stride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
2013-11-19 14:46:01 +01:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW && !COMPILE_TEMPLATE_AVX */
|
2013-11-19 14:57:52 +01:00
|
|
|
|
2013-12-14 02:22:44 +01:00
|
|
|
#if !COMPILE_TEMPLATE_AVX || HAVE_AVX_EXTERNAL
|
2013-11-19 22:36:27 +01:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW && (ARCH_X86_32 || COMPILE_TEMPLATE_SSE2) && COMPILE_TEMPLATE_MMXEXT == COMPILE_TEMPLATE_SSE2 && HAVE_YASM
|
2013-11-19 14:57:52 +01:00
|
|
|
void RENAME(ff_nv12ToUV)(uint8_t *dstU, uint8_t *dstV,
|
2014-01-21 22:19:16 +01:00
|
|
|
const uint8_t *unused,
|
|
|
|
|
const uint8_t *src1,
|
|
|
|
|
const uint8_t *src2,
|
|
|
|
|
int w,
|
2013-11-19 14:57:52 +01:00
|
|
|
uint32_t *unused2);
|
2013-11-19 14:57:52 +01:00
|
|
|
static void RENAME(deinterleaveBytes)(const uint8_t *src, uint8_t *dst1, uint8_t *dst2,
|
|
|
|
|
int width, int height, int srcStride,
|
|
|
|
|
int dst1Stride, int dst2Stride)
|
|
|
|
|
{
|
|
|
|
|
int h;
|
|
|
|
|
|
2013-11-19 14:57:52 +01:00
|
|
|
for (h = 0; h < height; h++) {
|
2013-11-19 14:57:52 +01:00
|
|
|
RENAME(ff_nv12ToUV)(dst1, dst2, NULL, src, NULL, width, NULL);
|
2013-11-19 14:57:52 +01:00
|
|
|
src += srcStride;
|
2013-11-19 14:57:52 +01:00
|
|
|
dst1 += dst1Stride;
|
|
|
|
|
dst2 += dst2Stride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
2015-06-23 06:14:42 +02:00
|
|
|
#if !COMPILE_TEMPLATE_SSE2
|
2013-11-19 14:57:52 +01:00
|
|
|
EMMS" \n\t"
|
2015-06-23 06:14:42 +02:00
|
|
|
#endif
|
2013-11-19 14:57:52 +01:00
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2013-12-14 02:22:44 +01:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AVX || HAVE_AVX_EXTERNAL */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_SSE2
|
|
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
static inline void RENAME(vu9_to_vu12)(const uint8_t *src1, const uint8_t *src2,
|
|
|
|
|
uint8_t *dst1, uint8_t *dst2,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int srcStride1, int srcStride2,
|
|
|
|
|
int dstStride1, int dstStride2)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
x86_reg x, y;
|
|
|
|
|
int w,h;
|
2011-05-25 05:23:39 +02:00
|
|
|
w=width/2; h=height/2;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
PREFETCH" %0 \n\t"
|
|
|
|
|
PREFETCH" %1 \n\t"
|
|
|
|
|
::"m"(*(src1+srcStride1)),"m"(*(src2+srcStride2)):"memory");
|
|
|
|
|
for (y=0;y<h;y++) {
|
|
|
|
|
const uint8_t* s1=src1+srcStride1*(y>>1);
|
|
|
|
|
uint8_t* d=dst1+dstStride1*y;
|
|
|
|
|
x=0;
|
|
|
|
|
for (;x<w-31;x+=32) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1,%2) \n\t"
|
|
|
|
|
"movq (%1,%2), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1,%2), %%mm2 \n\t"
|
|
|
|
|
"movq 16(%1,%2), %%mm4 \n\t"
|
|
|
|
|
"movq 24(%1,%2), %%mm6 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm5 \n\t"
|
|
|
|
|
"movq %%mm6, %%mm7 \n\t"
|
|
|
|
|
"punpcklbw %%mm0, %%mm0 \n\t"
|
|
|
|
|
"punpckhbw %%mm1, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm2, %%mm2 \n\t"
|
|
|
|
|
"punpckhbw %%mm3, %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm4, %%mm4 \n\t"
|
|
|
|
|
"punpckhbw %%mm5, %%mm5 \n\t"
|
|
|
|
|
"punpcklbw %%mm6, %%mm6 \n\t"
|
|
|
|
|
"punpckhbw %%mm7, %%mm7 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1, 8(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 16(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 24(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm4, 32(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm5, 40(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm6, 48(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm7, 56(%0,%2,2)"
|
|
|
|
|
:: "r"(d), "r"(s1), "r"(x)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
}
|
|
|
|
|
for (;x<w;x++) d[2*x]=d[2*x+1]=s1[x];
|
|
|
|
|
}
|
|
|
|
|
for (y=0;y<h;y++) {
|
|
|
|
|
const uint8_t* s2=src2+srcStride2*(y>>1);
|
|
|
|
|
uint8_t* d=dst2+dstStride2*y;
|
|
|
|
|
x=0;
|
|
|
|
|
for (;x<w-31;x+=32) {
|
|
|
|
|
__asm__ volatile(
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
PREFETCH" 32(%1,%2) \n\t"
|
|
|
|
|
"movq (%1,%2), %%mm0 \n\t"
|
|
|
|
|
"movq 8(%1,%2), %%mm2 \n\t"
|
|
|
|
|
"movq 16(%1,%2), %%mm4 \n\t"
|
|
|
|
|
"movq 24(%1,%2), %%mm6 \n\t"
|
2011-05-25 05:23:39 +02:00
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"movq %%mm4, %%mm5 \n\t"
|
|
|
|
|
"movq %%mm6, %%mm7 \n\t"
|
|
|
|
|
"punpcklbw %%mm0, %%mm0 \n\t"
|
|
|
|
|
"punpckhbw %%mm1, %%mm1 \n\t"
|
|
|
|
|
"punpcklbw %%mm2, %%mm2 \n\t"
|
|
|
|
|
"punpckhbw %%mm3, %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm4, %%mm4 \n\t"
|
|
|
|
|
"punpckhbw %%mm5, %%mm5 \n\t"
|
|
|
|
|
"punpcklbw %%mm6, %%mm6 \n\t"
|
|
|
|
|
"punpckhbw %%mm7, %%mm7 \n\t"
|
x86: swscale: fix fragile memory accesses
To access data at multiple fixed offsets from a base address, this
code uses a single "m" operand and code of the form "32%0", relying on
the memory operand instantiation having no displacement, giving a final
result of the form "32(%rax)". If the compiler uses a register and
displacement, e.g. "64(%rax)", the end result becomes "3264(%rax)",
which obviously does not work.
Replacing the "m" operands with "r" operands allows safe addition of a
displacement. In theory, multiple memory operands could use a shared
base register with different index registers, "(%rax,%rbx)", potentially
making more efficient use of registers. In the cases at hand, no such
sharing is possible since the addresses involved are entirely unrelated.
After this change, the code somewhat rudely accesses memory without
using a corresponding memory operand, which in some cases can lead to
unwanted "optimisations" of surrounding code. However, the original
code also accesses memory not covered by a memory operand, so this is
not adding any defect not already present. It is also hightly unlikely
that any such optimisations could be performed here since the memory
locations in questions are not accessed elsewhere in the same functions.
This fixes crashes with suncc.
Signed-off-by: Mans Rullgard <mans@mansr.com>
2012-08-13 01:53:05 +02:00
|
|
|
MOVNTQ" %%mm0, (%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1, 8(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2, 16(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 24(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm4, 32(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm5, 40(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm6, 48(%0,%2,2) \n\t"
|
|
|
|
|
MOVNTQ" %%mm7, 56(%0,%2,2)"
|
|
|
|
|
:: "r"(d), "r"(s2), "r"(x)
|
2011-05-25 05:23:39 +02:00
|
|
|
:"memory");
|
|
|
|
|
}
|
|
|
|
|
for (;x<w;x++) d[2*x]=d[2*x+1]=s2[x];
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static inline void RENAME(yvu9_to_yuy2)(const uint8_t *src1, const uint8_t *src2, const uint8_t *src3,
|
|
|
|
|
uint8_t *dst,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int srcStride1, int srcStride2,
|
|
|
|
|
int srcStride3, int dstStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
|
|
|
|
x86_reg x;
|
2011-05-26 18:57:26 +02:00
|
|
|
int y,w,h;
|
2011-05-25 05:23:39 +02:00
|
|
|
w=width/2; h=height;
|
|
|
|
|
for (y=0;y<h;y++) {
|
|
|
|
|
const uint8_t* yp=src1+srcStride1*y;
|
|
|
|
|
const uint8_t* up=src2+srcStride2*(y>>2);
|
|
|
|
|
const uint8_t* vp=src3+srcStride3*(y>>2);
|
|
|
|
|
uint8_t* d=dst+dstStride*y;
|
|
|
|
|
x=0;
|
|
|
|
|
for (;x<w-7;x+=8) {
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
PREFETCH" 32(%1, %0) \n\t"
|
|
|
|
|
PREFETCH" 32(%2, %0) \n\t"
|
|
|
|
|
PREFETCH" 32(%3, %0) \n\t"
|
|
|
|
|
"movq (%1, %0, 4), %%mm0 \n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
|
|
|
|
|
"movq (%2, %0), %%mm1 \n\t" /* U0U1U2U3U4U5U6U7 */
|
|
|
|
|
"movq (%3, %0), %%mm2 \n\t" /* V0V1V2V3V4V5V6V7 */
|
|
|
|
|
"movq %%mm0, %%mm3 \n\t" /* Y0Y1Y2Y3Y4Y5Y6Y7 */
|
|
|
|
|
"movq %%mm1, %%mm4 \n\t" /* U0U1U2U3U4U5U6U7 */
|
|
|
|
|
"movq %%mm2, %%mm5 \n\t" /* V0V1V2V3V4V5V6V7 */
|
|
|
|
|
"punpcklbw %%mm1, %%mm1 \n\t" /* U0U0 U1U1 U2U2 U3U3 */
|
|
|
|
|
"punpcklbw %%mm2, %%mm2 \n\t" /* V0V0 V1V1 V2V2 V3V3 */
|
|
|
|
|
"punpckhbw %%mm4, %%mm4 \n\t" /* U4U4 U5U5 U6U6 U7U7 */
|
|
|
|
|
"punpckhbw %%mm5, %%mm5 \n\t" /* V4V4 V5V5 V6V6 V7V7 */
|
|
|
|
|
|
|
|
|
|
"movq %%mm1, %%mm6 \n\t"
|
|
|
|
|
"punpcklbw %%mm2, %%mm1 \n\t" /* U0V0 U0V0 U1V1 U1V1*/
|
|
|
|
|
"punpcklbw %%mm1, %%mm0 \n\t" /* Y0U0 Y1V0 Y2U0 Y3V0*/
|
|
|
|
|
"punpckhbw %%mm1, %%mm3 \n\t" /* Y4U1 Y5V1 Y6U1 Y7V1*/
|
|
|
|
|
MOVNTQ" %%mm0, (%4, %0, 8) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 8(%4, %0, 8) \n\t"
|
|
|
|
|
|
|
|
|
|
"punpckhbw %%mm2, %%mm6 \n\t" /* U2V2 U2V2 U3V3 U3V3*/
|
|
|
|
|
"movq 8(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm6, %%mm0 \n\t" /* Y U2 Y V2 Y U2 Y V2*/
|
|
|
|
|
"punpckhbw %%mm6, %%mm3 \n\t" /* Y U3 Y V3 Y U3 Y V3*/
|
|
|
|
|
MOVNTQ" %%mm0, 16(%4, %0, 8) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 24(%4, %0, 8) \n\t"
|
|
|
|
|
|
|
|
|
|
"movq %%mm4, %%mm6 \n\t"
|
|
|
|
|
"movq 16(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm5, %%mm4 \n\t"
|
|
|
|
|
"punpcklbw %%mm4, %%mm0 \n\t" /* Y U4 Y V4 Y U4 Y V4*/
|
|
|
|
|
"punpckhbw %%mm4, %%mm3 \n\t" /* Y U5 Y V5 Y U5 Y V5*/
|
|
|
|
|
MOVNTQ" %%mm0, 32(%4, %0, 8) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 40(%4, %0, 8) \n\t"
|
|
|
|
|
|
|
|
|
|
"punpckhbw %%mm5, %%mm6 \n\t"
|
|
|
|
|
"movq 24(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm3 \n\t"
|
|
|
|
|
"punpcklbw %%mm6, %%mm0 \n\t" /* Y U6 Y V6 Y U6 Y V6*/
|
|
|
|
|
"punpckhbw %%mm6, %%mm3 \n\t" /* Y U7 Y V7 Y U7 Y V7*/
|
|
|
|
|
MOVNTQ" %%mm0, 48(%4, %0, 8) \n\t"
|
|
|
|
|
MOVNTQ" %%mm3, 56(%4, %0, 8) \n\t"
|
|
|
|
|
|
|
|
|
|
: "+r" (x)
|
|
|
|
|
: "r"(yp), "r" (up), "r"(vp), "r"(d)
|
|
|
|
|
:"memory");
|
|
|
|
|
}
|
|
|
|
|
for (; x<w; x++) {
|
2011-05-26 18:57:26 +02:00
|
|
|
const int x2 = x<<2;
|
2011-05-25 05:23:39 +02:00
|
|
|
d[8*x+0] = yp[x2];
|
|
|
|
|
d[8*x+1] = up[x];
|
|
|
|
|
d[8*x+2] = yp[x2+1];
|
|
|
|
|
d[8*x+3] = vp[x];
|
|
|
|
|
d[8*x+4] = yp[x2+2];
|
|
|
|
|
d[8*x+5] = up[x];
|
|
|
|
|
d[8*x+6] = yp[x2+3];
|
|
|
|
|
d[8*x+7] = vp[x];
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
static void RENAME(extract_even)(const uint8_t *src, uint8_t *dst, x86_reg count)
|
|
|
|
|
{
|
|
|
|
|
dst += count;
|
|
|
|
|
src += 2*count;
|
|
|
|
|
count= - count;
|
|
|
|
|
|
|
|
|
|
if(count <= -16) {
|
|
|
|
|
count += 15;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq -30(%1, %0, 2), %%mm0 \n\t"
|
|
|
|
|
"movq -22(%1, %0, 2), %%mm1 \n\t"
|
|
|
|
|
"movq -14(%1, %0, 2), %%mm2 \n\t"
|
|
|
|
|
"movq -6(%1, %0, 2), %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0,-15(%2, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2,- 7(%2, %0) \n\t"
|
|
|
|
|
"add $16, %0 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
: "+r"(count)
|
|
|
|
|
: "r"(src), "r"(dst)
|
|
|
|
|
);
|
|
|
|
|
count -= 15;
|
|
|
|
|
}
|
|
|
|
|
while(count<0) {
|
|
|
|
|
dst[count]= src[2*count];
|
|
|
|
|
count++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2014-07-13 05:24:50 +02:00
|
|
|
static void RENAME(extract_odd)(const uint8_t *src, uint8_t *dst, x86_reg count)
|
|
|
|
|
{
|
|
|
|
|
src ++;
|
|
|
|
|
dst += count;
|
|
|
|
|
src += 2*count;
|
|
|
|
|
count= - count;
|
|
|
|
|
|
|
|
|
|
if(count < -16) {
|
|
|
|
|
count += 16;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq -32(%1, %0, 2), %%mm0 \n\t"
|
|
|
|
|
"movq -24(%1, %0, 2), %%mm1 \n\t"
|
|
|
|
|
"movq -16(%1, %0, 2), %%mm2 \n\t"
|
|
|
|
|
"movq -8(%1, %0, 2), %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0,-16(%2, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm2,- 8(%2, %0) \n\t"
|
|
|
|
|
"add $16, %0 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
: "+r"(count)
|
|
|
|
|
: "r"(src), "r"(dst)
|
|
|
|
|
);
|
|
|
|
|
count -= 16;
|
|
|
|
|
}
|
|
|
|
|
while(count<0) {
|
|
|
|
|
dst[count]= src[2*count];
|
|
|
|
|
count++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
static void RENAME(extract_even2)(const uint8_t *src, uint8_t *dst0, uint8_t *dst1, x86_reg count)
|
|
|
|
|
{
|
|
|
|
|
dst0+= count;
|
|
|
|
|
dst1+= count;
|
|
|
|
|
src += 4*count;
|
|
|
|
|
count= - count;
|
|
|
|
|
if(count <= -8) {
|
|
|
|
|
count += 7;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq -28(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq -20(%1, %0, 4), %%mm1 \n\t"
|
|
|
|
|
"movq -12(%1, %0, 4), %%mm2 \n\t"
|
|
|
|
|
"movq -4(%1, %0, 4), %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlw $8, %%mm0 \n\t"
|
|
|
|
|
"psrlw $8, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm2, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm1 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0,- 7(%3, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1,- 7(%2, %0) \n\t"
|
|
|
|
|
"add $8, %0 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
: "+r"(count)
|
|
|
|
|
: "r"(src), "r"(dst0), "r"(dst1)
|
|
|
|
|
);
|
|
|
|
|
count -= 7;
|
|
|
|
|
}
|
|
|
|
|
while(count<0) {
|
|
|
|
|
dst0[count]= src[4*count+0];
|
|
|
|
|
dst1[count]= src[4*count+2];
|
|
|
|
|
count++;
|
|
|
|
|
}
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
static void RENAME(extract_even2avg)(const uint8_t *src0, const uint8_t *src1, uint8_t *dst0, uint8_t *dst1, x86_reg count)
|
|
|
|
|
{
|
|
|
|
|
dst0 += count;
|
|
|
|
|
dst1 += count;
|
|
|
|
|
src0 += 4*count;
|
|
|
|
|
src1 += 4*count;
|
|
|
|
|
count= - count;
|
|
|
|
|
#ifdef PAVGB
|
|
|
|
|
if(count <= -8) {
|
|
|
|
|
count += 7;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq -28(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq -20(%1, %0, 4), %%mm1 \n\t"
|
|
|
|
|
"movq -12(%1, %0, 4), %%mm2 \n\t"
|
|
|
|
|
"movq -4(%1, %0, 4), %%mm3 \n\t"
|
|
|
|
|
PAVGB" -28(%2, %0, 4), %%mm0 \n\t"
|
|
|
|
|
PAVGB" -20(%2, %0, 4), %%mm1 \n\t"
|
|
|
|
|
PAVGB" -12(%2, %0, 4), %%mm2 \n\t"
|
|
|
|
|
PAVGB" - 4(%2, %0, 4), %%mm3 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm0 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlw $8, %%mm0 \n\t"
|
|
|
|
|
"psrlw $8, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm2, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm1 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0,- 7(%4, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1,- 7(%3, %0) \n\t"
|
|
|
|
|
"add $8, %0 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
: "+r"(count)
|
|
|
|
|
: "r"(src0), "r"(src1), "r"(dst0), "r"(dst1)
|
|
|
|
|
);
|
|
|
|
|
count -= 7;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
while(count<0) {
|
|
|
|
|
dst0[count]= (src0[4*count+0]+src1[4*count+0])>>1;
|
|
|
|
|
dst1[count]= (src0[4*count+2]+src1[4*count+2])>>1;
|
|
|
|
|
count++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
static void RENAME(extract_odd2)(const uint8_t *src, uint8_t *dst0, uint8_t *dst1, x86_reg count)
|
|
|
|
|
{
|
|
|
|
|
dst0+= count;
|
|
|
|
|
dst1+= count;
|
|
|
|
|
src += 4*count;
|
|
|
|
|
count= - count;
|
|
|
|
|
if(count <= -8) {
|
|
|
|
|
count += 7;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq -28(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq -20(%1, %0, 4), %%mm1 \n\t"
|
|
|
|
|
"movq -12(%1, %0, 4), %%mm2 \n\t"
|
|
|
|
|
"movq -4(%1, %0, 4), %%mm3 \n\t"
|
|
|
|
|
"psrlw $8, %%mm0 \n\t"
|
|
|
|
|
"psrlw $8, %%mm1 \n\t"
|
|
|
|
|
"psrlw $8, %%mm2 \n\t"
|
|
|
|
|
"psrlw $8, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlw $8, %%mm0 \n\t"
|
|
|
|
|
"psrlw $8, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm2, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm1 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0,- 7(%3, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1,- 7(%2, %0) \n\t"
|
|
|
|
|
"add $8, %0 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
: "+r"(count)
|
|
|
|
|
: "r"(src), "r"(dst0), "r"(dst1)
|
|
|
|
|
);
|
|
|
|
|
count -= 7;
|
|
|
|
|
}
|
|
|
|
|
src++;
|
|
|
|
|
while(count<0) {
|
|
|
|
|
dst0[count]= src[4*count+0];
|
|
|
|
|
dst1[count]= src[4*count+2];
|
|
|
|
|
count++;
|
|
|
|
|
}
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
static void RENAME(extract_odd2avg)(const uint8_t *src0, const uint8_t *src1, uint8_t *dst0, uint8_t *dst1, x86_reg count)
|
|
|
|
|
{
|
|
|
|
|
dst0 += count;
|
|
|
|
|
dst1 += count;
|
|
|
|
|
src0 += 4*count;
|
|
|
|
|
src1 += 4*count;
|
|
|
|
|
count= - count;
|
|
|
|
|
#ifdef PAVGB
|
|
|
|
|
if(count <= -8) {
|
|
|
|
|
count += 7;
|
|
|
|
|
__asm__ volatile(
|
|
|
|
|
"pcmpeqw %%mm7, %%mm7 \n\t"
|
|
|
|
|
"psrlw $8, %%mm7 \n\t"
|
|
|
|
|
"1: \n\t"
|
|
|
|
|
"movq -28(%1, %0, 4), %%mm0 \n\t"
|
|
|
|
|
"movq -20(%1, %0, 4), %%mm1 \n\t"
|
|
|
|
|
"movq -12(%1, %0, 4), %%mm2 \n\t"
|
|
|
|
|
"movq -4(%1, %0, 4), %%mm3 \n\t"
|
|
|
|
|
PAVGB" -28(%2, %0, 4), %%mm0 \n\t"
|
|
|
|
|
PAVGB" -20(%2, %0, 4), %%mm1 \n\t"
|
|
|
|
|
PAVGB" -12(%2, %0, 4), %%mm2 \n\t"
|
|
|
|
|
PAVGB" - 4(%2, %0, 4), %%mm3 \n\t"
|
|
|
|
|
"psrlw $8, %%mm0 \n\t"
|
|
|
|
|
"psrlw $8, %%mm1 \n\t"
|
|
|
|
|
"psrlw $8, %%mm2 \n\t"
|
|
|
|
|
"psrlw $8, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm1, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm2 \n\t"
|
|
|
|
|
"movq %%mm0, %%mm1 \n\t"
|
|
|
|
|
"movq %%mm2, %%mm3 \n\t"
|
|
|
|
|
"psrlw $8, %%mm0 \n\t"
|
|
|
|
|
"psrlw $8, %%mm2 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm1 \n\t"
|
|
|
|
|
"pand %%mm7, %%mm3 \n\t"
|
|
|
|
|
"packuswb %%mm2, %%mm0 \n\t"
|
|
|
|
|
"packuswb %%mm3, %%mm1 \n\t"
|
|
|
|
|
MOVNTQ" %%mm0,- 7(%4, %0) \n\t"
|
|
|
|
|
MOVNTQ" %%mm1,- 7(%3, %0) \n\t"
|
|
|
|
|
"add $8, %0 \n\t"
|
|
|
|
|
" js 1b \n\t"
|
|
|
|
|
: "+r"(count)
|
|
|
|
|
: "r"(src0), "r"(src1), "r"(dst0), "r"(dst1)
|
|
|
|
|
);
|
|
|
|
|
count -= 7;
|
|
|
|
|
}
|
|
|
|
|
#endif
|
|
|
|
|
src0++;
|
|
|
|
|
src1++;
|
|
|
|
|
while(count<0) {
|
|
|
|
|
dst0[count]= (src0[4*count+0]+src1[4*count+0])>>1;
|
|
|
|
|
dst1[count]= (src0[4*count+2]+src1[4*count+2])>>1;
|
|
|
|
|
count++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void RENAME(yuyvtoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int srcStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2013-05-07 16:31:02 +02:00
|
|
|
const int chromWidth = FF_CEIL_RSHIFT(width, 1);
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
for (y=0; y<height; y++) {
|
|
|
|
|
RENAME(extract_even)(src, ydst, width);
|
|
|
|
|
if(y&1) {
|
|
|
|
|
RENAME(extract_odd2avg)(src-srcStride, src, udst, vdst, chromWidth);
|
|
|
|
|
udst+= chromStride;
|
|
|
|
|
vdst+= chromStride;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
src += srcStride;
|
|
|
|
|
ydst+= lumStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
static void RENAME(yuyvtoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int srcStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2013-05-07 16:31:02 +02:00
|
|
|
const int chromWidth = FF_CEIL_RSHIFT(width, 1);
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
for (y=0; y<height; y++) {
|
|
|
|
|
RENAME(extract_even)(src, ydst, width);
|
|
|
|
|
RENAME(extract_odd2)(src, udst, vdst, chromWidth);
|
|
|
|
|
|
|
|
|
|
src += srcStride;
|
|
|
|
|
ydst+= lumStride;
|
|
|
|
|
udst+= chromStride;
|
|
|
|
|
vdst+= chromStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
static void RENAME(uyvytoyuv420)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int srcStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2013-05-07 16:31:02 +02:00
|
|
|
const int chromWidth = FF_CEIL_RSHIFT(width, 1);
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
for (y=0; y<height; y++) {
|
2014-07-13 05:24:50 +02:00
|
|
|
RENAME(extract_odd)(src, ydst, width);
|
2011-05-25 05:23:39 +02:00
|
|
|
if(y&1) {
|
|
|
|
|
RENAME(extract_even2avg)(src-srcStride, src, udst, vdst, chromWidth);
|
|
|
|
|
udst+= chromStride;
|
|
|
|
|
vdst+= chromStride;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
src += srcStride;
|
|
|
|
|
ydst+= lumStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
|
|
|
|
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
static void RENAME(uyvytoyuv422)(uint8_t *ydst, uint8_t *udst, uint8_t *vdst, const uint8_t *src,
|
2011-05-26 18:57:26 +02:00
|
|
|
int width, int height,
|
|
|
|
|
int lumStride, int chromStride, int srcStride)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-26 18:57:26 +02:00
|
|
|
int y;
|
2013-05-07 16:31:02 +02:00
|
|
|
const int chromWidth = FF_CEIL_RSHIFT(width, 1);
|
2011-05-25 05:23:39 +02:00
|
|
|
|
|
|
|
|
for (y=0; y<height; y++) {
|
2014-07-13 05:24:50 +02:00
|
|
|
RENAME(extract_odd)(src, ydst, width);
|
2011-05-25 05:23:39 +02:00
|
|
|
RENAME(extract_even2)(src, udst, vdst, chromWidth);
|
|
|
|
|
|
|
|
|
|
src += srcStride;
|
|
|
|
|
ydst+= lumStride;
|
|
|
|
|
udst+= chromStride;
|
|
|
|
|
vdst+= chromStride;
|
|
|
|
|
}
|
|
|
|
|
__asm__(
|
|
|
|
|
EMMS" \n\t"
|
|
|
|
|
SFENCE" \n\t"
|
|
|
|
|
::: "memory"
|
|
|
|
|
);
|
|
|
|
|
}
|
2011-05-25 00:28:40 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
|
|
|
|
#endif /* !COMPILE_TEMPLATE_SSE2 */
|
2011-05-25 05:23:39 +02:00
|
|
|
|
2013-04-18 15:54:45 +02:00
|
|
|
static av_cold void RENAME(rgb2rgb_init)(void)
|
2011-05-25 05:23:39 +02:00
|
|
|
{
|
2011-05-25 00:28:40 +02:00
|
|
|
#if !COMPILE_TEMPLATE_SSE2
|
|
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 05:23:39 +02:00
|
|
|
rgb15to16 = RENAME(rgb15to16);
|
|
|
|
|
rgb15tobgr24 = RENAME(rgb15tobgr24);
|
|
|
|
|
rgb15to32 = RENAME(rgb15to32);
|
|
|
|
|
rgb16tobgr24 = RENAME(rgb16tobgr24);
|
|
|
|
|
rgb16to32 = RENAME(rgb16to32);
|
|
|
|
|
rgb16to15 = RENAME(rgb16to15);
|
|
|
|
|
rgb24tobgr16 = RENAME(rgb24tobgr16);
|
|
|
|
|
rgb24tobgr15 = RENAME(rgb24tobgr15);
|
|
|
|
|
rgb24tobgr32 = RENAME(rgb24tobgr32);
|
|
|
|
|
rgb32to16 = RENAME(rgb32to16);
|
|
|
|
|
rgb32to15 = RENAME(rgb32to15);
|
|
|
|
|
rgb32tobgr24 = RENAME(rgb32tobgr24);
|
|
|
|
|
rgb24to15 = RENAME(rgb24to15);
|
|
|
|
|
rgb24to16 = RENAME(rgb24to16);
|
|
|
|
|
rgb24tobgr24 = RENAME(rgb24tobgr24);
|
|
|
|
|
shuffle_bytes_2103 = RENAME(shuffle_bytes_2103);
|
|
|
|
|
rgb32tobgr16 = RENAME(rgb32tobgr16);
|
|
|
|
|
rgb32tobgr15 = RENAME(rgb32tobgr15);
|
|
|
|
|
yv12toyuy2 = RENAME(yv12toyuy2);
|
|
|
|
|
yv12touyvy = RENAME(yv12touyvy);
|
|
|
|
|
yuv422ptoyuy2 = RENAME(yuv422ptoyuy2);
|
|
|
|
|
yuv422ptouyvy = RENAME(yuv422ptouyvy);
|
|
|
|
|
yuy2toyv12 = RENAME(yuy2toyv12);
|
|
|
|
|
vu9_to_vu12 = RENAME(vu9_to_vu12);
|
|
|
|
|
yvu9_to_yuy2 = RENAME(yvu9_to_yuy2);
|
|
|
|
|
uyvytoyuv422 = RENAME(uyvytoyuv422);
|
|
|
|
|
yuyvtoyuv422 = RENAME(yuyvtoyuv422);
|
2012-08-16 14:35:41 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW */
|
2011-05-25 00:28:40 +02:00
|
|
|
|
2012-07-08 18:42:12 +02:00
|
|
|
#if COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW
|
2011-05-25 00:28:40 +02:00
|
|
|
planar2x = RENAME(planar2x);
|
2012-07-08 18:42:12 +02:00
|
|
|
#endif /* COMPILE_TEMPLATE_MMXEXT || COMPILE_TEMPLATE_AMD3DNOW */
|
2013-04-18 12:47:16 +02:00
|
|
|
#if HAVE_7REGS
|
2013-04-15 03:01:04 +02:00
|
|
|
ff_rgb24toyv12 = RENAME(rgb24toyv12);
|
2013-04-18 12:47:16 +02:00
|
|
|
#endif /* HAVE_7REGS */
|
2011-05-25 00:28:40 +02:00
|
|
|
|
|
|
|
|
yuyvtoyuv420 = RENAME(yuyvtoyuv420);
|
|
|
|
|
uyvytoyuv420 = RENAME(uyvytoyuv420);
|
2012-08-16 14:35:41 +02:00
|
|
|
#endif /* !COMPILE_TEMPLATE_SSE2 */
|
2011-05-25 00:28:40 +02:00
|
|
|
|
2013-11-19 14:46:01 +01:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW && !COMPILE_TEMPLATE_AVX
|
2011-05-25 00:28:40 +02:00
|
|
|
interleaveBytes = RENAME(interleaveBytes);
|
2013-11-19 14:46:01 +01:00
|
|
|
#endif /* !COMPILE_TEMPLATE_AMD3DNOW && !COMPILE_TEMPLATE_AVX */
|
2013-12-16 03:38:13 +01:00
|
|
|
#if !COMPILE_TEMPLATE_AVX || HAVE_AVX_EXTERNAL
|
2013-11-19 22:36:27 +01:00
|
|
|
#if !COMPILE_TEMPLATE_AMD3DNOW && (ARCH_X86_32 || COMPILE_TEMPLATE_SSE2) && COMPILE_TEMPLATE_MMXEXT == COMPILE_TEMPLATE_SSE2 && HAVE_YASM
|
2013-11-19 14:57:52 +01:00
|
|
|
deinterleaveBytes = RENAME(deinterleaveBytes);
|
|
|
|
|
#endif
|
2013-12-16 03:38:13 +01:00
|
|
|
#endif
|
2011-05-25 05:23:39 +02:00
|
|
|
}
|