coffee/libyuv
1
0
Fork 0
mirror of https://chromium.googlesource.com/libyuv/libyuv synced 2025-12-07 01:06:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

Move scale row functions to scale_win etc

Browse Source 
BUG=none
TEST=untested
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/4509005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@879 16f28f9a-4ce2-e073-06de-1de4eb20be90
This commit is contained in:
fbarchard@google.com 2013-11-28 01:16:15 +00:00
parent 9cd5bc0ceb
commit dbe4814361
15 changed files with 3418 additions and 3256 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Android.mk
View File

@ -29,6 +29,7 @@ LOCAL_SRC_FILES := \
source/scale_argb.cc \
source/scale_common.cc \
source/scale_mips.cc \
source/scale_posix.cc \
source/video_common.cc
# TODO(fbarchard): Enable mjpeg encoder.

2
README.chromium
View File

@ -1,6 +1,6 @@
Name: libyuv
URL: http://code.google.com/p/libyuv/
Version: 877
Version: 879
License: BSD
License File: LICENSE

228
include/libyuv/scale_row.h
View File

@ -18,6 +18,43 @@ namespace libyuv {
extern "C" {
#endif
// The following are available on all x86 platforms:
#if !defined(LIBYUV_DISABLE_X86) && \
(defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
#define HAS_SCALEROWDOWN2_SSE2
#define HAS_SCALEROWDOWN4_SSE2
#define HAS_SCALEROWDOWN34_SSSE3
#define HAS_SCALEROWDOWN38_SSSE3
#define HAS_SCALEADDROWS_SSE2
#define HAS_SCALEFILTERCOLS_SSSE3
#define HAS_SCALECOLSUP2_SSE2
#define HAS_SCALEARGBROWDOWN2_SSE2
#define HAS_SCALEARGBROWDOWNEVEN_SSE2
#define HAS_SCALEARGBCOLS_SSE2
#define HAS_SCALEARGBFILTERCOLS_SSSE3
#define HAS_SCALEARGBCOLSUP2_SSE2
#endif
// The following are available on Neon platforms:
#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
(defined(__ARM_NEON__) || defined(LIBYUV_NEON))
#define HAS_SCALEROWDOWN2_NEON
#define HAS_SCALEROWDOWN4_NEON
#define HAS_SCALEROWDOWN34_NEON
#define HAS_SCALEROWDOWN38_NEON
#define HAS_SCALEARGBROWDOWNEVEN_NEON
#define HAS_SCALEARGBROWDOWN2_NEON
#endif
// The following are available on Mips platforms:
#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
defined(__mips__)
#define HAS_SCALEROWDOWN2_MIPS_DSPR2
#define HAS_SCALEROWDOWN4_MIPS_DSPR2
#define HAS_SCALEROWDOWN34_MIPS_DSPR2
#define HAS_SCALEROWDOWN38_MIPS_DSPR2
#endif
// Scale ARGB vertically with bilinear interpolation.
void ScalePlaneVertical(int src_height,
int dst_width, int dst_height,
@ -26,6 +63,197 @@ void ScalePlaneVertical(int src_height,
int x, int y, int dy,
int bpp, FilterMode filtering);
void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* d, int dst_width);
void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* d, int dst_width);
void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int x, int dx);
void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int, int);
void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int x, int dx);
void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint16* dst_ptr, int src_width, int src_height);
void ScaleARGBRowDown2_C(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t /* src_stride */,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx);
void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int, int);
void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx);
void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t,
uint8* dst_ptr, int dst_width);
void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint16* dst_ptr, int src_width,
int src_height);
void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int x, int dx);
void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int /* x */, int /* dx */);
void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
ptrdiff_t src_stride,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx);
void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx);
void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int /* x */, int /* dx */);
// Row functions.
void ScaleARGBRowDownEven_NEON(const uint8* src_argb, int src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, int src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
// ScaleRowDown2Box also used by planar functions
// NEON downscalers with interpolation.
// Note - not static due to reuse in convert for 444 to 420.
void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst_ptr, int dst_width);
void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
// Down scale from 4 to 3 pixels. Use the neon multilane read/write
// to load up the every 4th pixel into a 4 different registers.
// Point samples 32 pixels to 24 pixels.
void ScaleRowDown34_NEON(const uint8* src_ptr,
ptrdiff_t /* src_stride */,
uint8* dst_ptr, int dst_width);
void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
// 32 -> 12
void ScaleRowDown38_NEON(const uint8* src_ptr,
ptrdiff_t /* src_stride */,
uint8* dst_ptr, int dst_width);
// 32x3 -> 12x1
void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
// 32x2 -> 12x1
void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* d, int dst_width);
void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* d, int dst_width);
void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width);
#ifdef __cplusplus
} // extern "C"
} // namespace libyuv

2
include/libyuv/version.h
View File

@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT
#define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 877
#define LIBYUV_VERSION 879
#endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT

5
libyuv.gyp
View File

@ -74,6 +74,7 @@
'include/libyuv/row.h',
'include/libyuv/scale.h',
'include/libyuv/scale_argb.h',
'include/libyuv/scale_row.h',
'include/libyuv/version.h',
'include/libyuv/video_common.h',
@ -107,10 +108,12 @@
'source/row_win.cc',
'source/scale.cc',
'source/scale_argb.cc',
'source/scale_argb_neon.cc',
'source/scale_argb_neon.cc', # Deprecated.
'source/scale_common.cc',
'source/scale_mips.cc',
'source/scale_neon.cc',
'source/scale_posix.cc',
'source/scale_win.cc',
'source/video_common.cc',
],
},

1
linux.mk
View File

@ -28,6 +28,7 @@ LOCAL_OBJ_FILES := \
source/scale_argb.o \
source/scale_common.o \
source/scale_mips.o \
source/scale_posix.o \
source/video_common.o
.cc.o:

5
source/row_mips.cc
View File

@ -15,6 +15,9 @@ namespace libyuv {
extern "C" {
#endif
// The following are available on Mips platforms:
#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__)
#ifdef HAS_COPYROW_MIPS
void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
__asm__ __volatile__ (
@ -968,6 +971,8 @@ void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
}
#endif // __mips_dsp_rev >= 2
#endif // defined(__mips__)
#ifdef __cplusplus
} // extern "C"
} // namespace libyuv

4
source/row_posix.cc
View File

@ -10,14 +10,12 @@
#include "libyuv/row.h"
#include "libyuv/basic_types.h"
#ifdef __cplusplus
namespace libyuv {
extern "C" {
#endif
// This module is for GCC x86 and x64
// This module is for GCC x86 and x64.
#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
// TODO(nfullagar): For Native Client: When new toolchain becomes available,

2119
source/scale.cc
View File

File diff suppressed because it is too large Load Diff

988
source/scale_argb.cc
View File

@ -27,994 +27,6 @@ static __inline int Abs(int v) {
return v >= 0 ? v : -v;
}
// ARGB scaling uses bilinear or point, but not box filter.
#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
(defined(__ARM_NEON__) || defined(LIBYUV_NEON))
#define HAS_SCALEARGBROWDOWNEVEN_NEON
#define HAS_SCALEARGBROWDOWN2_NEON
void ScaleARGBRowDownEven_NEON(const uint8* src_argb, int src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, int src_stride,
int src_stepx,
uint8* dst_argb, int dst_width);
void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width);
void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width);
#endif
#if !defined(LIBYUV_DISABLE_X86) && \
defined(_M_IX86) && defined(_MSC_VER)
#define HAS_SCALEARGBROWDOWN2_SSE2
// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
__declspec(naked) __declspec(align(16))
static void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
__asm {
mov eax, [esp + 4] // src_argb
// src_stride ignored
mov edx, [esp + 12] // dst_argb
mov ecx, [esp + 16] // dst_width
align 16
wloop:
movdqa xmm0, [eax]
movdqa xmm1, [eax + 16]
lea eax, [eax + 32]
shufps xmm0, xmm1, 0xdd
sub ecx, 4
movdqa [edx], xmm0
lea edx, [edx + 16]
jg wloop
ret
}
}
// Blends 8x1 rectangle to 4x1.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
__declspec(naked) __declspec(align(16))
static void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
__asm {
mov eax, [esp + 4] // src_argb
// src_stride ignored
mov edx, [esp + 12] // dst_argb
mov ecx, [esp + 16] // dst_width
align 16
wloop:
movdqa xmm0, [eax]
movdqa xmm1, [eax + 16]
lea eax, [eax + 32]
movdqa xmm2, xmm0
shufps xmm0, xmm1, 0x88 // even pixels
shufps xmm2, xmm1, 0xdd // odd pixels
pavgb xmm0, xmm2
sub ecx, 4
movdqa [edx], xmm0
lea edx, [edx + 16]
jg wloop
ret
}
}
// Blends 8x2 rectangle to 4x1.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
__declspec(naked) __declspec(align(16))
static void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) {
__asm {
push esi
mov eax, [esp + 4 + 4] // src_argb
mov esi, [esp + 4 + 8] // src_stride
mov edx, [esp + 4 + 12] // dst_argb
mov ecx, [esp + 4 + 16] // dst_width
align 16
wloop:
movdqa xmm0, [eax]
movdqa xmm1, [eax + 16]
movdqa xmm2, [eax + esi]
movdqa xmm3, [eax + esi + 16]
lea eax, [eax + 32]
pavgb xmm0, xmm2 // average rows
pavgb xmm1, xmm3
movdqa xmm2, xmm0 // average columns (8 to 4 pixels)
shufps xmm0, xmm1, 0x88 // even pixels
shufps xmm2, xmm1, 0xdd // odd pixels
pavgb xmm0, xmm2
sub ecx, 4
movdqa [edx], xmm0
lea edx, [edx + 16]
jg wloop
pop esi
ret
}
}
#define HAS_SCALEARGBROWDOWNEVEN_SSE2
// Reads 4 pixels at a time.
// Alignment requirement: dst_argb 16 byte aligned.
__declspec(naked) __declspec(align(16))
void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
__asm {
push ebx
push edi
mov eax, [esp + 8 + 4] // src_argb
// src_stride ignored
mov ebx, [esp + 8 + 12] // src_stepx
mov edx, [esp + 8 + 16] // dst_argb
mov ecx, [esp + 8 + 20] // dst_width
lea ebx, [ebx * 4]
lea edi, [ebx + ebx * 2]
align 16
wloop:
movd xmm0, [eax]
movd xmm1, [eax + ebx]
punpckldq xmm0, xmm1
movd xmm2, [eax + ebx * 2]
movd xmm3, [eax + edi]
lea eax, [eax + ebx * 4]
punpckldq xmm2, xmm3
punpcklqdq xmm0, xmm2
sub ecx, 4
movdqa [edx], xmm0
lea edx, [edx + 16]
jg wloop
pop edi
pop ebx
ret
}
}
// Blends four 2x2 to 4x1.
// Alignment requirement: dst_argb 16 byte aligned.
__declspec(naked) __declspec(align(16))
static void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
__asm {
push ebx
push esi
push edi
mov eax, [esp + 12 + 4] // src_argb
mov esi, [esp + 12 + 8] // src_stride
mov ebx, [esp + 12 + 12] // src_stepx
mov edx, [esp + 12 + 16] // dst_argb
mov ecx, [esp + 12 + 20] // dst_width
lea esi, [eax + esi] // row1 pointer
lea ebx, [ebx * 4]
lea edi, [ebx + ebx * 2]
align 16
wloop:
movq xmm0, qword ptr [eax] // row0 4 pairs
movhps xmm0, qword ptr [eax + ebx]
movq xmm1, qword ptr [eax + ebx * 2]
movhps xmm1, qword ptr [eax + edi]
lea eax, [eax + ebx * 4]
movq xmm2, qword ptr [esi] // row1 4 pairs
movhps xmm2, qword ptr [esi + ebx]
movq xmm3, qword ptr [esi + ebx * 2]
movhps xmm3, qword ptr [esi + edi]
lea esi, [esi + ebx * 4]
pavgb xmm0, xmm2 // average rows
pavgb xmm1, xmm3
movdqa xmm2, xmm0 // average columns (8 to 4 pixels)
shufps xmm0, xmm1, 0x88 // even pixels
shufps xmm2, xmm1, 0xdd // odd pixels
pavgb xmm0, xmm2
sub ecx, 4
movdqa [edx], xmm0
lea edx, [edx + 16]
jg wloop
pop edi
pop esi
pop ebx
ret
}
}
// Column scaling unfiltered. SSE2 version.
#define HAS_SCALEARGBCOLS_SSE2
__declspec(naked) __declspec(align(16))
void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
__asm {
push edi
push esi
mov edi, [esp + 8 + 4] // dst_argb
mov esi, [esp + 8 + 8] // src_argb
mov ecx, [esp + 8 + 12] // dst_width
movd xmm2, [esp + 8 + 16] // x
movd xmm3, [esp + 8 + 20] // dx
pshufd xmm2, xmm2, 0 // x0 x0 x0 x0
pshufd xmm0, xmm3, 0x11 // dx 0 dx 0
paddd xmm2, xmm0
paddd xmm3, xmm3 // 0, 0, 0, dx * 2
pshufd xmm0, xmm3, 0x05 // dx * 2, dx * 2, 0, 0
paddd xmm2, xmm0 // x3 x2 x1 x0
paddd xmm3, xmm3 // 0, 0, 0, dx * 4
pshufd xmm3, xmm3, 0 // dx * 4, dx * 4, dx * 4, dx * 4
pextrw eax, xmm2, 1 // get x0 integer.
pextrw edx, xmm2, 3 // get x1 integer.
cmp ecx, 0
jle xloop99
sub ecx, 4
jl xloop49
// 4 Pixel loop.
align 4
xloop4:
movd xmm0, [esi + eax * 4] // 1 source x0 pixels
movd xmm1, [esi + edx * 4] // 1 source x1 pixels
pextrw eax, xmm2, 5 // get x2 integer.
pextrw edx, xmm2, 7 // get x3 integer.
paddd xmm2, xmm3 // x += dx
punpckldq xmm0, xmm1 // x0 x1
movd xmm1, [esi + eax * 4] // 1 source x2 pixels
movd xmm4, [esi + edx * 4] // 1 source x3 pixels
pextrw eax, xmm2, 1 // get x0 integer. next iteration.
pextrw edx, xmm2, 3 // get x1 integer. next iteration.
punpckldq xmm1, xmm4 // x2 x3
punpcklqdq xmm0, xmm1 // x0 x1 x2 x3
sub ecx, 4 // 4 pixels
movdqu [edi], xmm0
lea edi, [edi + 16]
jge xloop4
align 4
xloop49:
test ecx, 2
je xloop29
// 2 Pixels.
movd xmm0, [esi + eax * 4] // 1 source x0 pixels
movd xmm1, [esi + edx * 4] // 1 source x1 pixels
pextrw eax, xmm2, 5 // get x2 integer.
punpckldq xmm0, xmm1 // x0 x1
movq qword ptr [edi], xmm0
lea edi, [edi + 8]
xloop29:
test ecx, 1
je xloop99
// 1 Pixels.
movd xmm0, [esi + eax * 4] // 1 source x2 pixels
movd dword ptr [edi], xmm0
align 4
xloop99:
pop esi
pop edi
ret
}
}
// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version.
// TODO(fbarchard): Port to Neon
// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
static uvec8 kShuffleColARGB = {
0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u, // bbggrraa 1st pixel
8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u // bbggrraa 2nd pixel
};
// Shuffle table for duplicating 2 fractions into 8 bytes each
static uvec8 kShuffleFractions = {
0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
};
#define HAS_SCALEARGBFILTERCOLS_SSSE3
__declspec(naked) __declspec(align(16))
static void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
__asm {
push esi
push edi
mov edi, [esp + 8 + 4] // dst_argb
mov esi, [esp + 8 + 8] // src_argb
mov ecx, [esp + 8 + 12] // dst_width
movd xmm2, [esp + 8 + 16] // x
movd xmm3, [esp + 8 + 20] // dx
movdqa xmm4, kShuffleColARGB
movdqa xmm5, kShuffleFractions
pcmpeqb xmm6, xmm6 // generate 0x007f for inverting fraction.
psrlw xmm6, 9
pextrw eax, xmm2, 1 // get x0 integer. preroll
sub ecx, 2
jl xloop29
movdqa xmm0, xmm2 // x1 = x0 + dx
paddd xmm0, xmm3
punpckldq xmm2, xmm0 // x0 x1
punpckldq xmm3, xmm3 // dx dx
paddd xmm3, xmm3 // dx * 2, dx * 2
pextrw edx, xmm2, 3 // get x1 integer. preroll
// 2 Pixel loop.
align 4
xloop2:
movdqa xmm1, xmm2 // x0, x1 fractions.
paddd xmm2, xmm3 // x += dx
movq xmm0, qword ptr [esi + eax * 4] // 2 source x0 pixels
psrlw xmm1, 9 // 7 bit fractions.
movhps xmm0, qword ptr [esi + edx * 4] // 2 source x1 pixels
pshufb xmm1, xmm5 // 0000000011111111
pshufb xmm0, xmm4 // arrange pixels into pairs
pxor xmm1, xmm6 // 0..7f and 7f..0
pmaddubsw xmm0, xmm1 // argb_argb 16 bit, 2 pixels.
pextrw eax, xmm2, 1 // get x0 integer. next iteration.
pextrw edx, xmm2, 3 // get x1 integer. next iteration.
psrlw xmm0, 7 // argb 8.7 fixed point to low 8 bits.
packuswb xmm0, xmm0 // argb_argb 8 bits, 2 pixels.
movq qword ptr [edi], xmm0
lea edi, [edi + 8]
sub ecx, 2 // 2 pixels
jge xloop2
align 4
xloop29:
add ecx, 2 - 1
jl xloop99
// 1 pixel remainder
psrlw xmm2, 9 // 7 bit fractions.
movq xmm0, qword ptr [esi + eax * 4] // 2 source x0 pixels
pshufb xmm2, xmm5 // 00000000
pshufb xmm0, xmm4 // arrange pixels into pairs
pxor xmm2, xmm6 // 0..7f and 7f..0
pmaddubsw xmm0, xmm2 // argb 16 bit, 1 pixel.
psrlw xmm0, 7
packuswb xmm0, xmm0 // argb 8 bits, 1 pixel.
movd [edi], xmm0
align 16
xloop99:
pop edi
pop esi
ret
}
}
// Reads 4 pixels, duplicates them and writes 8 pixels.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
#define HAS_SCALEARGBCOLSUP2_SSE2
__declspec(naked) __declspec(align(16))
void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int /* x */, int /* dx */) {
__asm {
mov edx, [esp + 4] // dst_argb
mov eax, [esp + 8] // src_argb
mov ecx, [esp + 12] // dst_width
align 16
wloop:
movdqa xmm0, [eax]
lea eax, [eax + 16]
movdqa xmm1, xmm0
punpckldq xmm0, xmm0
punpckhdq xmm1, xmm1
sub ecx, 8
movdqa [edx], xmm0
movdqa [edx + 16], xmm1
lea edx, [edx + 32]
jg wloop
ret
}
}
#elif !defined(LIBYUV_DISABLE_X86) && \
(defined(__x86_64__) || defined(__i386__))
// TODO(nfullagar): For Native Client: When new toolchain becomes available,
// take advantage of bundle lock / unlock feature. This will reduce the amount
// of manual bundle alignment done below, and bundle alignment could even be
// moved into each macro that doesn't use %%nacl: such as MEMOPREG.
#if defined(__native_client__) && defined(__x86_64__)
#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
#define MEMLEA(offset, base) #offset "(%q" #base ")"
#define MEMLEA3(offset, index, scale) \
#offset "(,%q" #index "," #scale ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%q" #base ",%q" #index "," #scale ")"
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " (%%r15,%%r14),%%" #reg "\n"
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
"lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
#opcode " %%" #reg ",(%%r15,%%r14)\n"
#define BUNDLEALIGN ".p2align 5 \n"
#else
#define MEMACCESS(base) "(%" #base ")"
#define MEMACCESS2(offset, base) #offset "(%" #base ")"
#define MEMLEA(offset, base) #offset "(%" #base ")"
#define MEMLEA3(offset, index, scale) \
#offset "(,%" #index "," #scale ")"
#define MEMLEA4(offset, base, index, scale) \
#offset "(%" #base ",%" #index "," #scale ")"
#define MEMOPREG(opcode, offset, base, index, scale, reg) \
#opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
#opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
#define BUNDLEALIGN
#endif
// GCC versions of row functions are verbatim conversions from Visual C,
// with some additional macro injection for Native Client (see row_posix.cc
// for more details.)
// Generated using gcc disassembly on Visual C object file:
// objdump -D yuvscaler.obj >yuvscaler.txt
#define HAS_SCALEARGBROWDOWN2_SSE2
static void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
asm volatile (
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
"lea " MEMLEA(0x20,0) ",%0 \n"
"shufps $0xdd,%%xmm1,%%xmm0 \n"
"sub $0x4,%2 \n"
"movdqa %%xmm0," MEMACCESS(1) " \n"
"lea " MEMLEA(0x10,1) ",%1 \n"
"jg 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_argb), // %1
"+r"(dst_width) // %2
:
: "memory", "cc"
#if defined(__SSE2__)
, "xmm0", "xmm1"
#endif
);
}
static void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
asm volatile (
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
"lea " MEMLEA(0x20,0) ",%0 \n"
"movdqa %%xmm0,%%xmm2 \n"
"shufps $0x88,%%xmm1,%%xmm0 \n"
"shufps $0xdd,%%xmm1,%%xmm2 \n"
"pavgb %%xmm2,%%xmm0 \n"
"sub $0x4,%2 \n"
"movdqa %%xmm0," MEMACCESS(1) " \n"
"lea " MEMLEA(0x10,1) ",%1 \n"
"jg 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_argb), // %1
"+r"(dst_width) // %2
:
: "memory", "cc"
#if defined(__SSE2__)
, "xmm0", "xmm1"
#endif
);
}
static void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) {
asm volatile (
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movdqa " MEMACCESS(0) ",%%xmm0 \n"
"movdqa " MEMACCESS2(0x10,0) ",%%xmm1 \n"
BUNDLEALIGN
MEMOPREG(movdqa,0x00,0,3,1,xmm2) // movdqa (%0,%3,1),%%xmm2
MEMOPREG(movdqa,0x10,0,3,1,xmm3) // movdqa 0x10(%0,%3,1),%%xmm3
"lea " MEMLEA(0x20,0) ",%0 \n"
"pavgb %%xmm2,%%xmm0 \n"
"pavgb %%xmm3,%%xmm1 \n"
"movdqa %%xmm0,%%xmm2 \n"
"shufps $0x88,%%xmm1,%%xmm0 \n"
"shufps $0xdd,%%xmm1,%%xmm2 \n"
"pavgb %%xmm2,%%xmm0 \n"
"sub $0x4,%2 \n"
"movdqa %%xmm0," MEMACCESS(1) " \n"
"lea " MEMLEA(0x10,1) ",%1 \n"
"jg 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_argb), // %1
"+r"(dst_width) // %2
: "r"(static_cast<intptr_t>(src_stride)) // %3
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3"
#endif
);
}
#define HAS_SCALEARGBROWDOWNEVEN_SSE2
// Reads 4 pixels at a time.
// Alignment requirement: dst_argb 16 byte aligned.
void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
intptr_t src_stepx_x4 = static_cast<intptr_t>(src_stepx);
intptr_t src_stepx_x12 = 0;
asm volatile (
"lea " MEMLEA3(0x00,1,4) ",%1 \n"
"lea " MEMLEA4(0x00,1,1,2) ",%4 \n"
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movd " MEMACCESS(0) ",%%xmm0 \n"
MEMOPREG(movd,0x00,0,1,1,xmm1) // movd (%0,%1,1),%%xmm1
"punpckldq %%xmm1,%%xmm0 \n"
BUNDLEALIGN
MEMOPREG(movd,0x00,0,1,2,xmm2) // movd (%0,%1,2),%%xmm2
MEMOPREG(movd,0x00,0,4,1,xmm3) // movd (%0,%4,1),%%xmm3
"lea " MEMLEA4(0x00,0,1,4) ",%0 \n"
"punpckldq %%xmm3,%%xmm2 \n"
"punpcklqdq %%xmm2,%%xmm0 \n"
"sub $0x4,%3 \n"
"movdqa %%xmm0," MEMACCESS(2) " \n"
"lea " MEMLEA(0x10,2) ",%2 \n"
"jg 1b \n"
: "+r"(src_argb), // %0
"+r"(src_stepx_x4), // %1
"+r"(dst_argb), // %2
"+r"(dst_width), // %3
"+r"(src_stepx_x12) // %4
:
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3"
#endif
);
}
// Blends four 2x2 to 4x1.
// Alignment requirement: dst_argb 16 byte aligned.
static void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
ptrdiff_t src_stride, int src_stepx,
uint8* dst_argb, int dst_width) {
intptr_t src_stepx_x4 = static_cast<intptr_t>(src_stepx);
intptr_t src_stepx_x12 = 0;
intptr_t row1 = static_cast<intptr_t>(src_stride);
asm volatile (
"lea " MEMLEA3(0x00,1,4) ",%1 \n"
"lea " MEMLEA4(0x00,1,1,2) ",%4 \n"
"lea " MEMLEA4(0x00,0,5,1) ",%5 \n"
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movq " MEMACCESS(0) ",%%xmm0 \n"
MEMOPREG(movhps,0x00,0,1,1,xmm0) // movhps (%0,%1,1),%%xmm0
MEMOPREG(movq,0x00,0,1,2,xmm1) // movq (%0,%1,2),%%xmm1
BUNDLEALIGN
MEMOPREG(movhps,0x00,0,4,1,xmm1) // movhps (%0,%4,1),%%xmm1
"lea " MEMLEA4(0x00,0,1,4) ",%0 \n"
"movq " MEMACCESS(5) ",%%xmm2 \n"
BUNDLEALIGN
MEMOPREG(movhps,0x00,5,1,1,xmm2) // movhps (%5,%1,1),%%xmm2
MEMOPREG(movq,0x00,5,1,2,xmm3) // movq (%5,%1,2),%%xmm3
MEMOPREG(movhps,0x00,5,4,1,xmm3) // movhps (%5,%4,1),%%xmm3
"lea " MEMLEA4(0x00,5,1,4) ",%5 \n"
"pavgb %%xmm2,%%xmm0 \n"
"pavgb %%xmm3,%%xmm1 \n"
"movdqa %%xmm0,%%xmm2 \n"
"shufps $0x88,%%xmm1,%%xmm0 \n"
"shufps $0xdd,%%xmm1,%%xmm2 \n"
"pavgb %%xmm2,%%xmm0 \n"
"sub $0x4,%3 \n"
"movdqa %%xmm0," MEMACCESS(2) " \n"
"lea " MEMLEA(0x10,2) ",%2 \n"
"jg 1b \n"
: "+r"(src_argb), // %0
"+r"(src_stepx_x4), // %1
"+r"(dst_argb), // %2
"+rm"(dst_width), // %3
"+r"(src_stepx_x12), // %4
"+r"(row1) // %5
:
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3"
#endif
);
}
#define HAS_SCALEARGBCOLS_SSE2
void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
intptr_t x0 = 0, x1 = 0;
asm volatile (
"movd %5,%%xmm2 \n"
"movd %6,%%xmm3 \n"
"pshufd $0x0,%%xmm2,%%xmm2 \n"
"pshufd $0x11,%%xmm3,%%xmm0 \n"
"paddd %%xmm0,%%xmm2 \n"
"paddd %%xmm3,%%xmm3 \n"
"pshufd $0x5,%%xmm3,%%xmm0 \n"
"paddd %%xmm0,%%xmm2 \n"
"paddd %%xmm3,%%xmm3 \n"
"pshufd $0x0,%%xmm3,%%xmm3 \n"
"pextrw $0x1,%%xmm2,%k0 \n"
"pextrw $0x3,%%xmm2,%k1 \n"
"cmp $0x0,%4 \n"
"jl 99f \n"
"sub $0x4,%4 \n"
"jl 49f \n"
".p2align 2 \n"
BUNDLEALIGN
"40: \n"
MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0
MEMOPREG(movd,0x00,3,1,4,xmm1) // movd (%3,%1,4),%%xmm1
"pextrw $0x5,%%xmm2,%k0 \n"
"pextrw $0x7,%%xmm2,%k1 \n"
"paddd %%xmm3,%%xmm2 \n"
"punpckldq %%xmm1,%%xmm0 \n"
MEMOPREG(movd,0x00,3,0,4,xmm1) // movd (%3,%0,4),%%xmm1
MEMOPREG(movd,0x00,3,1,4,xmm4) // movd (%3,%1,4),%%xmm4
"pextrw $0x1,%%xmm2,%k0 \n"
"pextrw $0x3,%%xmm2,%k1 \n"
"punpckldq %%xmm4,%%xmm1 \n"
"punpcklqdq %%xmm1,%%xmm0 \n"
"sub $0x4,%4 \n"
"movdqu %%xmm0," MEMACCESS(2) " \n"
"lea " MEMLEA(0x10,2) ",%2 \n"
"jge 40b \n"
"49: \n"
"test $0x2,%4 \n"
"je 29f \n"
BUNDLEALIGN
MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0
MEMOPREG(movd,0x00,3,1,4,xmm1) // movd (%3,%1,4),%%xmm1
"pextrw $0x5,%%xmm2,%k0 \n"
"punpckldq %%xmm1,%%xmm0 \n"
"movq %%xmm0," MEMACCESS(2) " \n"
"lea " MEMLEA(0x8,2) ",%2 \n"
"29: \n"
"test $0x1,%4 \n"
"je 99f \n"
MEMOPREG(movd,0x00,3,0,4,xmm0) // movd (%3,%0,4),%%xmm0
"movd %%xmm0," MEMACCESS(2) " \n"
"99: \n"
: "+a"(x0), // %0
"+d"(x1), // %1
"+r"(dst_argb), // %2
"+r"(src_argb), // %3
"+r"(dst_width) // %4
: "rm"(x), // %5
"rm"(dx) // %6
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
#endif
);
}
// Reads 4 pixels, duplicates them and writes 8 pixels.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
#define HAS_SCALEARGBCOLSUP2_SSE2
void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
int dst_width, int /* x */, int /* dx */) {
asm volatile (
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movdqa " MEMACCESS(1) ",%%xmm0 \n"
"lea " MEMLEA(0x10,1) ",%1 \n"
"movdqa %%xmm0,%%xmm1 \n"
"punpckldq %%xmm0,%%xmm0 \n"
"punpckhdq %%xmm1,%%xmm1 \n"
"sub $0x8,%2 \n"
"movdqa %%xmm0," MEMACCESS(0) " \n"
"movdqa %%xmm1," MEMACCESS2(0x10,0) " \n"
"lea " MEMLEA(0x20,0) ",%0 \n"
"jg 1b \n"
: "+r"(dst_argb), // %0
"+r"(src_argb), // %1
"+r"(dst_width) // %2
:
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1"
#endif
);
}
// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
static uvec8 kShuffleColARGB = {
0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u, // bbggrraa 1st pixel
8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u // bbggrraa 2nd pixel
};
// Shuffle table for duplicating 2 fractions into 8 bytes each
static uvec8 kShuffleFractions = {
0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
};
// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
#define HAS_SCALEARGBFILTERCOLS_SSSE3
static void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
intptr_t x0 = 0, x1 = 0;
asm volatile (
"movdqa %0,%%xmm4 \n"
"movdqa %1,%%xmm5 \n"
:
: "m"(kShuffleColARGB), // %0
"m"(kShuffleFractions) // %1
);
asm volatile (
"movd %5,%%xmm2 \n"
"movd %6,%%xmm3 \n"
"pcmpeqb %%xmm6,%%xmm6 \n"
"psrlw $0x9,%%xmm6 \n"
"pextrw $0x1,%%xmm2,%k3 \n"
"sub $0x2,%2 \n"
"jl 29f \n"
"movdqa %%xmm2,%%xmm0 \n"
"paddd %%xmm3,%%xmm0 \n"
"punpckldq %%xmm0,%%xmm2 \n"
"punpckldq %%xmm3,%%xmm3 \n"
"paddd %%xmm3,%%xmm3 \n"
"pextrw $0x3,%%xmm2,%k4 \n"
".p2align 2 \n"
BUNDLEALIGN
"2: \n"
"movdqa %%xmm2,%%xmm1 \n"
"paddd %%xmm3,%%xmm2 \n"
MEMOPREG(movq,0x00,1,3,4,xmm0) // movq (%1,%3,4),%%xmm0
"psrlw $0x9,%%xmm1 \n"
BUNDLEALIGN
MEMOPREG(movhps,0x00,1,4,4,xmm0) // movhps (%1,%4,4),%%xmm0
"pshufb %%xmm5,%%xmm1 \n"
"pshufb %%xmm4,%%xmm0 \n"
"pxor %%xmm6,%%xmm1 \n"
"pmaddubsw %%xmm1,%%xmm0 \n"
"psrlw $0x7,%%xmm0 \n"
"pextrw $0x1,%%xmm2,%k3 \n"
"pextrw $0x3,%%xmm2,%k4 \n"
"packuswb %%xmm0,%%xmm0 \n"
"movq %%xmm0," MEMACCESS(0) " \n"
"lea " MEMLEA(0x8,0) ",%0 \n"
"sub $0x2,%2 \n"
"jge 2b \n"
".p2align 2 \n"
BUNDLEALIGN
"29: \n"
"add $0x1,%2 \n"
"jl 99f \n"
"psrlw $0x9,%%xmm2 \n"
BUNDLEALIGN
MEMOPREG(movq,0x00,1,3,4,xmm0) // movq (%1,%3,4),%%xmm0
"pshufb %%xmm5,%%xmm2 \n"
"pshufb %%xmm4,%%xmm0 \n"
"pxor %%xmm6,%%xmm2 \n"
"pmaddubsw %%xmm2,%%xmm0 \n"
"psrlw $0x7,%%xmm0 \n"
"packuswb %%xmm0,%%xmm0 \n"
"movd %%xmm0," MEMACCESS(0) " \n"
".p2align 4 \n"
"99: \n"
: "+r"(dst_argb), // %0
"+r"(src_argb), // %1
"+rm"(dst_width), // %2
"+r"(x0), // %3
"+r"(x1) // %4
: "rm"(x), // %5
"rm"(dx) // %6
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
#endif
);
}
#endif // defined(__x86_64__) || defined(__i386__)
static void ScaleARGBRowDown2_C(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int x = 0; x < dst_width - 1; x += 2) {
dst[0] = src[1];
dst[1] = src[3];
src += 4;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[1];
}
}
static void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) {
dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
src_argb += 8;
dst_argb += 4;
}
}
static void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) {
dst_argb[0] = (src_argb[0] + src_argb[4] +
src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
dst_argb[1] = (src_argb[1] + src_argb[5] +
src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
dst_argb[2] = (src_argb[2] + src_argb[6] +
src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
dst_argb[3] = (src_argb[3] + src_argb[7] +
src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
src_argb += 8;
dst_argb += 4;
}
}
void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t /* src_stride */,
int src_stepx,
uint8* dst_argb, int dst_width) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int x = 0; x < dst_width - 1; x += 2) {
dst[0] = src[0];
dst[1] = src[src_stepx];
src += src_stepx * 2;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[0];
}
}
static void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) {
dst_argb[0] = (src_argb[0] + src_argb[4] +
src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
dst_argb[1] = (src_argb[1] + src_argb[5] +
src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
dst_argb[2] = (src_argb[2] + src_argb[6] +
src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
dst_argb[3] = (src_argb[3] + src_argb[7] +
src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
src_argb += src_stepx * 4;
dst_argb += 4;
}
}
// Scales a single row of pixels using point sampling.
void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int j = 0; j < dst_width - 1; j += 2) {
dst[0] = src[x >> 16];
x += dx;
dst[1] = src[x >> 16];
x += dx;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[x >> 16];
}
}
// Scales a single row of pixels up by 2x using point sampling.
void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int, int) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int j = 0; j < dst_width - 1; j += 2) {
dst[1] = dst[0] = src[0];
src += 1;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[0];
}
}
// Mimics SSSE3 blender
#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
#define BLENDERC(a, b, f, s) static_cast<uint32>( \
BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
#define BLENDER(a, b, f) \
BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
static void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int j = 0; j < dst_width - 1; j += 2) {
int xi = x >> 16;
int xf = (x >> 9) & 0x7f;
uint32 a = src[xi];
uint32 b = src[xi + 1];
dst[0] = BLENDER(a, b, xf);
x += dx;
xi = x >> 16;
xf = (x >> 9) & 0x7f;
a = src[xi];
b = src[xi + 1];
dst[1] = BLENDER(a, b, xf);
x += dx;
dst += 2;
}
if (dst_width & 1) {
int xi = x >> 16;
int xf = (x >> 9) & 0x7f;
uint32 a = src[xi];
uint32 b = src[xi + 1];
dst[0] = BLENDER(a, b, xf);
}
}
// ScaleARGB ARGB, 1/2
// This is an optimized version for scaling down a ARGB to 1/2 of
// its original size.

142
source/scale_argb_neon.cc
View File

@ -1,142 +1,2 @@
/*
* Copyright 2012 The LibYuv Project Authors. All rights reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
// This file is deprecated.
#include "libyuv/basic_types.h"
#include "libyuv/row.h"
#ifdef __cplusplus
namespace libyuv {
extern "C" {
#endif
// This module is for GCC Neon
#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width) {
asm volatile (
"1: \n"
// load even pixels into q0, odd into q1
"vld2.32 {q0, q1}, [%0]! \n"
"vld2.32 {q2, q3}, [%0]! \n"
"subs %2, %2, #8 \n" // 8 processed per loop
"vst1.8 {q1}, [%1]! \n" // store odd pixels
"vst1.8 {q3}, [%1]! \n"
"bgt 1b \n"
: "+r"(src_ptr), // %0
"+r"(dst), // %1
"+r"(dst_width) // %2
:
: "memory", "cc", "q0", "q1", "q2", "q3" // Clobber List
);
}
void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width) {
asm volatile (
// change the stride to row 2 pointer
"add %1, %1, %0 \n"
"1: \n"
"vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels.
"vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels.
"subs %3, %3, #8 \n" // 8 processed per loop.
"vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts.
"vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts.
"vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts.
"vpaddl.u8 q3, q3 \n" // A 16 bytes -> 8 shorts.
"vld4.8 {d16, d18, d20, d22}, [%1]! \n" // load 8 more ARGB pixels.
"vld4.8 {d17, d19, d21, d23}, [%1]! \n" // load last 8 ARGB pixels.
"vpadal.u8 q0, q8 \n" // B 16 bytes -> 8 shorts.
"vpadal.u8 q1, q9 \n" // G 16 bytes -> 8 shorts.
"vpadal.u8 q2, q10 \n" // R 16 bytes -> 8 shorts.
"vpadal.u8 q3, q11 \n" // A 16 bytes -> 8 shorts.
"vrshrn.u16 d0, q0, #2 \n" // downshift, round and pack
"vrshrn.u16 d1, q1, #2 \n"
"vrshrn.u16 d2, q2, #2 \n"
"vrshrn.u16 d3, q3, #2 \n"
"vst4.8 {d0, d1, d2, d3}, [%2]! \n"
"bgt 1b \n"
: "+r"(src_ptr), // %0
"+r"(src_stride), // %1
"+r"(dst), // %2
"+r"(dst_width) // %3
:
: "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
);
}
// Reads 4 pixels at a time.
// Alignment requirement: src_argb 4 byte aligned.
void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t, int src_stepx,
uint8* dst_argb, int dst_width) {
asm volatile (
"mov r12, %3, lsl #2 \n"
".p2align 2 \n"
"1: \n"
"vld1.32 {d0[0]}, [%0], r12 \n"
"vld1.32 {d0[1]}, [%0], r12 \n"
"vld1.32 {d1[0]}, [%0], r12 \n"
"vld1.32 {d1[1]}, [%0], r12 \n"
"subs %2, %2, #4 \n" // 4 pixels per loop.
"vst1.8 {q0}, [%1]! \n"
"bgt 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_argb), // %1
"+r"(dst_width) // %2
: "r"(src_stepx) // %3
: "memory", "cc", "r12", "q0"
);
}
// Reads 4 pixels at a time.
// Alignment requirement: src_argb 4 byte aligned.
void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
asm volatile (
"mov r12, %4, lsl #2 \n"
"add %1, %1, %0 \n"
".p2align 2 \n"
"1: \n"
"vld1.8 {d0}, [%0], r12 \n" // Read 4 2x2 blocks -> 2x1
"vld1.8 {d1}, [%1], r12 \n"
"vld1.8 {d2}, [%0], r12 \n"
"vld1.8 {d3}, [%1], r12 \n"
"vld1.8 {d4}, [%0], r12 \n"
"vld1.8 {d5}, [%1], r12 \n"
"vld1.8 {d6}, [%0], r12 \n"
"vld1.8 {d7}, [%1], r12 \n"
"vaddl.u8 q0, d0, d1 \n"
"vaddl.u8 q1, d2, d3 \n"
"vaddl.u8 q2, d4, d5 \n"
"vaddl.u8 q3, d6, d7 \n"
"vswp.8 d1, d2 \n" // ab_cd -> ac_bd
"vswp.8 d5, d6 \n" // ef_gh -> eg_fh
"vadd.u16 q0, q0, q1 \n" // (a+b)_(c+d)
"vadd.u16 q2, q2, q3 \n" // (e+f)_(g+h)
"vrshrn.u16 d0, q0, #2 \n" // first 2 pixels.
"vrshrn.u16 d1, q2, #2 \n" // next 2 pixels.
"subs %3, %3, #4 \n" // 4 pixels per loop.
"vst1.8 {q0}, [%2]! \n"
"bgt 1b \n"
: "+r"(src_argb), // %0
"+r"(src_stride), // %1
"+r"(dst_argb), // %2
"+r"(dst_width) // %3
: "r"(src_stepx) // %4
: "memory", "cc", "r12", "q0", "q1", "q2", "q3"
);
}
#endif // __ARM_NEON__
#ifdef __cplusplus
} // extern "C"
} // namespace libyuv
#endif

438
source/scale_common.cc
View File

@ -23,6 +23,444 @@ namespace libyuv {
extern "C" {
#endif
// CPU agnostic row functions
void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width) {
uint8* dend = dst + dst_width - 1;
do {
dst[0] = src_ptr[1];
dst[1] = src_ptr[3];
dst += 2;
src_ptr += 4;
} while (dst < dend);
if (dst_width & 1) {
dst[0] = src_ptr[1];
}
}
void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width) {
const uint8* s = src_ptr;
uint8* dend = dst + dst_width - 1;
do {
dst[0] = (s[0] + s[1] + 1) >> 1;
dst[1] = (s[2] + s[3] + 1) >> 1;
dst += 2;
s += 4;
} while (dst < dend);
if (dst_width & 1) {
dst[0] = (s[0] + s[1] + 1) >> 1;
}
}
void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width) {
const uint8* s = src_ptr;
const uint8* t = src_ptr + src_stride;
uint8* dend = dst + dst_width - 1;
do {
dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
dst += 2;
s += 4;
t += 4;
} while (dst < dend);
if (dst_width & 1) {
dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
}
}
void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width) {
uint8* dend = dst + dst_width - 1;
do {
dst[0] = src_ptr[2];
dst[1] = src_ptr[6];
dst += 2;
src_ptr += 8;
} while (dst < dend);
if (dst_width & 1) {
dst[0] = src_ptr[2];
}
}
void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width) {
intptr_t stride = src_stride;
uint8* dend = dst + dst_width - 1;
do {
dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
src_ptr[stride + 0] + src_ptr[stride + 1] +
src_ptr[stride + 2] + src_ptr[stride + 3] +
src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
8) >> 4;
dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
src_ptr[stride + 4] + src_ptr[stride + 5] +
src_ptr[stride + 6] + src_ptr[stride + 7] +
src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
8) >> 4;
dst += 2;
src_ptr += 8;
} while (dst < dend);
if (dst_width & 1) {
dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
src_ptr[stride + 0] + src_ptr[stride + 1] +
src_ptr[stride + 2] + src_ptr[stride + 3] +
src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
8) >> 4;
}
}
void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width) {
assert((dst_width % 3 == 0) && (dst_width > 0));
uint8* dend = dst + dst_width;
do {
dst[0] = src_ptr[0];
dst[1] = src_ptr[1];
dst[2] = src_ptr[3];
dst += 3;
src_ptr += 4;
} while (dst < dend);
}
// Filter rows 0 and 1 together, 3 : 1
void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* d, int dst_width) {
assert((dst_width % 3 == 0) && (dst_width > 0));
const uint8* s = src_ptr;
const uint8* t = src_ptr + src_stride;
uint8* dend = d + dst_width;
do {
uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
d[0] = (a0 * 3 + b0 + 2) >> 2;
d[1] = (a1 * 3 + b1 + 2) >> 2;
d[2] = (a2 * 3 + b2 + 2) >> 2;
d += 3;
s += 4;
t += 4;
} while (d < dend);
}
// Filter rows 1 and 2 together, 1 : 1
void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* d, int dst_width) {
assert((dst_width % 3 == 0) && (dst_width > 0));
const uint8* s = src_ptr;
const uint8* t = src_ptr + src_stride;
uint8* dend = d + dst_width;
do {
uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
d[0] = (a0 + b0 + 1) >> 1;
d[1] = (a1 + b1 + 1) >> 1;
d[2] = (a2 + b2 + 1) >> 1;
d += 3;
s += 4;
t += 4;
} while (d < dend);
}
// Scales a single row of pixels using point sampling.
void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int x, int dx) {
for (int j = 0; j < dst_width - 1; j += 2) {
dst_ptr[0] = src_ptr[x >> 16];
x += dx;
dst_ptr[1] = src_ptr[x >> 16];
x += dx;
dst_ptr += 2;
}
if (dst_width & 1) {
dst_ptr[0] = src_ptr[x >> 16];
}
}
// Scales a single row of pixels up by 2x using point sampling.
void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int, int) {
for (int j = 0; j < dst_width - 1; j += 2) {
dst_ptr[1] = dst_ptr[0] = src_ptr[0];
src_ptr += 1;
dst_ptr += 2;
}
if (dst_width & 1) {
dst_ptr[0] = src_ptr[0];
}
}
// (1-f)a + fb can be replaced with a + f(b-a)
#define BLENDER(a, b, f) (static_cast<int>(a) + \
((f) * (static_cast<int>(b) - static_cast<int>(a)) >> 16))
void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
int dst_width, int x, int dx) {
for (int j = 0; j < dst_width - 1; j += 2) {
int xi = x >> 16;
int a = src_ptr[xi];
int b = src_ptr[xi + 1];
dst_ptr[0] = BLENDER(a, b, x & 0xffff);
x += dx;
xi = x >> 16;
a = src_ptr[xi];
b = src_ptr[xi + 1];
dst_ptr[1] = BLENDER(a, b, x & 0xffff);
x += dx;
dst_ptr += 2;
}
if (dst_width & 1) {
int xi = x >> 16;
int a = src_ptr[xi];
int b = src_ptr[xi + 1];
dst_ptr[0] = BLENDER(a, b, x & 0xffff);
}
}
#undef BLENDER
void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width) {
assert(dst_width % 3 == 0);
for (int x = 0; x < dst_width; x += 3) {
dst[0] = src_ptr[0];
dst[1] = src_ptr[3];
dst[2] = src_ptr[6];
dst += 3;
src_ptr += 8;
}
}
// 8x3 -> 3x1
void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) {
assert((dst_width % 3 == 0) && (dst_width > 0));
intptr_t stride = src_stride;
for (int i = 0; i < dst_width; i += 3) {
dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
src_ptr[stride + 0] + src_ptr[stride + 1] +
src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
(65536 / 9) >> 16;
dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
src_ptr[stride + 3] + src_ptr[stride + 4] +
src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
(65536 / 9) >> 16;
dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
src_ptr[stride + 6] + src_ptr[stride + 7] +
src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
(65536 / 6) >> 16;
src_ptr += 8;
dst_ptr += 3;
}
}
// 8x2 -> 3x1
void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) {
assert((dst_width % 3 == 0) && (dst_width > 0));
intptr_t stride = src_stride;
for (int i = 0; i < dst_width; i += 3) {
dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
src_ptr[stride + 0] + src_ptr[stride + 1] +
src_ptr[stride + 2]) * (65536 / 6) >> 16;
dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
src_ptr[stride + 3] + src_ptr[stride + 4] +
src_ptr[stride + 5]) * (65536 / 6) >> 16;
dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
src_ptr[stride + 6] + src_ptr[stride + 7]) *
(65536 / 4) >> 16;
src_ptr += 8;
dst_ptr += 3;
}
}
void ScaleAddRows_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint16* dst_ptr, int src_width, int src_height) {
assert(src_width > 0);
assert(src_height > 0);
for (int x = 0; x < src_width; ++x) {
const uint8* s = src_ptr + x;
unsigned int sum = 0u;
for (int y = 0; y < src_height; ++y) {
sum += s[0];
s += src_stride;
}
// TODO(fbarchard): Consider limitting height to 256 to avoid overflow.
dst_ptr[x] = sum < 65535u ? sum : 65535u;
}
}
void ScaleARGBRowDown2_C(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int x = 0; x < dst_width - 1; x += 2) {
dst[0] = src[1];
dst[1] = src[3];
src += 4;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[1];
}
}
void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
ptrdiff_t /* src_stride */,
uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) {
dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
src_argb += 8;
dst_argb += 4;
}
}
void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) {
dst_argb[0] = (src_argb[0] + src_argb[4] +
src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
dst_argb[1] = (src_argb[1] + src_argb[5] +
src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
dst_argb[2] = (src_argb[2] + src_argb[6] +
src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
dst_argb[3] = (src_argb[3] + src_argb[7] +
src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
src_argb += 8;
dst_argb += 4;
}
}
void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t /* src_stride */,
int src_stepx,
uint8* dst_argb, int dst_width) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int x = 0; x < dst_width - 1; x += 2) {
dst[0] = src[0];
dst[1] = src[src_stepx];
src += src_stepx * 2;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[0];
}
}
void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) {
dst_argb[0] = (src_argb[0] + src_argb[4] +
src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
dst_argb[1] = (src_argb[1] + src_argb[5] +
src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
dst_argb[2] = (src_argb[2] + src_argb[6] +
src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
dst_argb[3] = (src_argb[3] + src_argb[7] +
src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
src_argb += src_stepx * 4;
dst_argb += 4;
}
}
// Scales a single row of pixels using point sampling.
void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int j = 0; j < dst_width - 1; j += 2) {
dst[0] = src[x >> 16];
x += dx;
dst[1] = src[x >> 16];
x += dx;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[x >> 16];
}
}
// Scales a single row of pixels up by 2x using point sampling.
void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int, int) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int j = 0; j < dst_width - 1; j += 2) {
dst[1] = dst[0] = src[0];
src += 1;
dst += 2;
}
if (dst_width & 1) {
dst[0] = src[0];
}
}
// Mimics SSSE3 blender
#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
#define BLENDERC(a, b, f, s) static_cast<uint32>( \
BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
#define BLENDER(a, b, f) \
BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
int dst_width, int x, int dx) {
const uint32* src = reinterpret_cast<const uint32*>(src_argb);
uint32* dst = reinterpret_cast<uint32*>(dst_argb);
for (int j = 0; j < dst_width - 1; j += 2) {
int xi = x >> 16;
int xf = (x >> 9) & 0x7f;
uint32 a = src[xi];
uint32 b = src[xi + 1];
dst[0] = BLENDER(a, b, xf);
x += dx;
xi = x >> 16;
xf = (x >> 9) & 0x7f;
a = src[xi];
b = src[xi + 1];
dst[1] = BLENDER(a, b, xf);
x += dx;
dst += 2;
}
if (dst_width & 1) {
int xi = x >> 16;
int xf = (x >> 9) & 0x7f;
uint32 a = src[xi];
uint32 b = src[xi + 1];
dst[0] = BLENDER(a, b, xf);
}
}
#undef BLENDER1
#undef BLENDERC
#undef BLENDER
// Scale plane vertically with bilinear interpolation.
void ScalePlaneVertical(int src_height,
int dst_width, int dst_height,

120
source/scale_neon.cc
View File

@ -8,7 +8,6 @@
* be found in the AUTHORS file in the root of the source tree.
*/
#include "libyuv/basic_types.h"
#include "libyuv/row.h"
#ifdef __cplusplus
@ -16,7 +15,7 @@ namespace libyuv {
extern "C" {
#endif
// This module is for GCC Neon
// This module is for GCC Neon.
#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)
// NEON downscalers with interpolation.
@ -546,6 +545,123 @@ void ScaleFilterRows_NEON(uint8* dst_ptr,
: "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
);
}
void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t /* src_stride */,
uint8* dst, int dst_width) {
asm volatile (
"1: \n"
// load even pixels into q0, odd into q1
"vld2.32 {q0, q1}, [%0]! \n"
"vld2.32 {q2, q3}, [%0]! \n"
"subs %2, %2, #8 \n" // 8 processed per loop
"vst1.8 {q1}, [%1]! \n" // store odd pixels
"vst1.8 {q3}, [%1]! \n"
"bgt 1b \n"
: "+r"(src_ptr), // %0
"+r"(dst), // %1
"+r"(dst_width) // %2
:
: "memory", "cc", "q0", "q1", "q2", "q3" // Clobber List
);
}
void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width) {
asm volatile (
// change the stride to row 2 pointer
"add %1, %1, %0 \n"
"1: \n"
"vld4.8 {d0, d2, d4, d6}, [%0]! \n" // load 8 ARGB pixels.
"vld4.8 {d1, d3, d5, d7}, [%0]! \n" // load next 8 ARGB pixels.
"subs %3, %3, #8 \n" // 8 processed per loop.
"vpaddl.u8 q0, q0 \n" // B 16 bytes -> 8 shorts.
"vpaddl.u8 q1, q1 \n" // G 16 bytes -> 8 shorts.
"vpaddl.u8 q2, q2 \n" // R 16 bytes -> 8 shorts.
"vpaddl.u8 q3, q3 \n" // A 16 bytes -> 8 shorts.
"vld4.8 {d16, d18, d20, d22}, [%1]! \n" // load 8 more ARGB pixels.
"vld4.8 {d17, d19, d21, d23}, [%1]! \n" // load last 8 ARGB pixels.
"vpadal.u8 q0, q8 \n" // B 16 bytes -> 8 shorts.
"vpadal.u8 q1, q9 \n" // G 16 bytes -> 8 shorts.
"vpadal.u8 q2, q10 \n" // R 16 bytes -> 8 shorts.
"vpadal.u8 q3, q11 \n" // A 16 bytes -> 8 shorts.
"vrshrn.u16 d0, q0, #2 \n" // downshift, round and pack
"vrshrn.u16 d1, q1, #2 \n"
"vrshrn.u16 d2, q2, #2 \n"
"vrshrn.u16 d3, q3, #2 \n"
"vst4.8 {d0, d1, d2, d3}, [%2]! \n"
"bgt 1b \n"
: "+r"(src_ptr), // %0
"+r"(src_stride), // %1
"+r"(dst), // %2
"+r"(dst_width) // %3
:
: "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
);
}
// Reads 4 pixels at a time.
// Alignment requirement: src_argb 4 byte aligned.
void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t, int src_stepx,
uint8* dst_argb, int dst_width) {
asm volatile (
"mov r12, %3, lsl #2 \n"
".p2align 2 \n"
"1: \n"
"vld1.32 {d0[0]}, [%0], r12 \n"
"vld1.32 {d0[1]}, [%0], r12 \n"
"vld1.32 {d1[0]}, [%0], r12 \n"
"vld1.32 {d1[1]}, [%0], r12 \n"
"subs %2, %2, #4 \n" // 4 pixels per loop.
"vst1.8 {q0}, [%1]! \n"
"bgt 1b \n"
: "+r"(src_argb), // %0
"+r"(dst_argb), // %1
"+r"(dst_width) // %2
: "r"(src_stepx) // %3
: "memory", "cc", "r12", "q0"
);
}
// Reads 4 pixels at a time.
// Alignment requirement: src_argb 4 byte aligned.
void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
int src_stepx,
uint8* dst_argb, int dst_width) {
asm volatile (
"mov r12, %4, lsl #2 \n"
"add %1, %1, %0 \n"
".p2align 2 \n"
"1: \n"
"vld1.8 {d0}, [%0], r12 \n" // Read 4 2x2 blocks -> 2x1
"vld1.8 {d1}, [%1], r12 \n"
"vld1.8 {d2}, [%0], r12 \n"
"vld1.8 {d3}, [%1], r12 \n"
"vld1.8 {d4}, [%0], r12 \n"
"vld1.8 {d5}, [%1], r12 \n"
"vld1.8 {d6}, [%0], r12 \n"
"vld1.8 {d7}, [%1], r12 \n"
"vaddl.u8 q0, d0, d1 \n"
"vaddl.u8 q1, d2, d3 \n"
"vaddl.u8 q2, d4, d5 \n"
"vaddl.u8 q3, d6, d7 \n"
"vswp.8 d1, d2 \n" // ab_cd -> ac_bd
"vswp.8 d5, d6 \n" // ef_gh -> eg_fh
"vadd.u16 q0, q0, q1 \n" // (a+b)_(c+d)
"vadd.u16 q2, q2, q3 \n" // (e+f)_(g+h)
"vrshrn.u16 d0, q0, #2 \n" // first 2 pixels.
"vrshrn.u16 d1, q2, #2 \n" // next 2 pixels.
"subs %3, %3, #4 \n" // 4 pixels per loop.
"vst1.8 {q0}, [%2]! \n"
"bgt 1b \n"
: "+r"(src_argb), // %0
"+r"(src_stride), // %1
"+r"(dst_argb), // %2
"+r"(dst_width) // %3
: "r"(src_stepx) // %4
: "memory", "cc", "r12", "q0", "q1", "q2", "q3"
);
}
#endif // __ARM_NEON__
#ifdef __cplusplus

1337
source/scale_posix.cc Normal file
View File

File diff suppressed because it is too large Load Diff

1282
source/scale_win.cc Normal file
View File

File diff suppressed because it is too large Load Diff