Skip to content

L
libyuv
Commit 788f7570 authored by fbarchard@google.com's avatar fbarchard@google.com
Browse files
Options

Linear interpolation. 

BUG=none
TEST=*Linear*
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/3689004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@848 16f28f9a-4ce2-e073-06de-1de4eb20be90
parent c2a889eb
Hide whitespace changes
Inline Side-by-side
Showing with 345 additions and 85 deletions
README.chromium
View file @ 788f7570
Name: libyuv Name: libyuv
URL: http://code.google.com/p/libyuv/ URL: http://code.google.com/p/libyuv/
Version: 847 Version: 848
License: BSD License: BSD
License File: LICENSE License File: LICENSE
......
include/libyuv/scale.h
View file @ 788f7570
...@@ -22,7 +22,8 @@ extern "C" { ...@@ -22,7 +22,8 @@ extern "C" {
enum FilterMode { enum FilterMode {
kFilterNone = 0, // Point sample; Fastest. kFilterNone = 0, // Point sample; Fastest.
kFilterBilinear = 1, // Faster than box, but lower quality scaling down. kFilterBilinear = 1, // Faster than box, but lower quality scaling down.
kFilterBox = 2 // Highest quality. kFilterBox = 2, // Highest quality.
kFilterLinear = 3 // Faster than bilinear, slower than None.
}; };
// Scale a YUV plane. // Scale a YUV plane.
......
include/libyuv/version.h
View file @ 788f7570
...@@ -11,6 +11,6 @@ ...@@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT #ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT
#define INCLUDE_LIBYUV_VERSION_H_ #define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 847 #define LIBYUV_VERSION 848
#endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT #endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT
source/scale.cc
View file @ 788f7570
...@@ -37,19 +37,7 @@ static __inline int Half(int v) { ...@@ -37,19 +37,7 @@ static __inline int Half(int v) {
// Note: Some SSE2 reference manuals // Note: Some SSE2 reference manuals
// cpuvol1.pdf agner_instruction_tables.pdf 253666.pdf 253667.pdf // cpuvol1.pdf agner_instruction_tables.pdf 253666.pdf 253667.pdf
// Set the following flag to true to revert to only // ScaleRowDown2Box also used by planar functions
// using the reference implementation ScalePlaneBox(), and
// NOT the optimized versions. Useful for debugging and
// when comparing the quality of the resulting YUV planes
// as produced by the optimized and non-optimized versions.
static bool use_reference_impl_ = false;
LIBYUV_API
void SetUseReferenceImpl(bool use) {
use_reference_impl_ = use;
}
// ScaleRowDown2Int also used by planar functions
// NEON downscalers with interpolation. // NEON downscalers with interpolation.
#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \ #if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
...@@ -208,6 +196,44 @@ static void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -208,6 +196,44 @@ static void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
} }
} }
// Blends 32x1 rectangle to 16x1.
// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
__declspec(naked) __declspec(align(16))
void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) {
__asm {
mov eax, [esp + 4] // src_ptr
// src_stride
mov edx, [esp + 12] // dst_ptr
mov ecx, [esp + 16] // dst_width
pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff
psrlw xmm5, 8
align 16
wloop:
movdqa xmm0, [eax]
movdqa xmm1, [eax + 16]
lea eax, [eax + 32]
movdqa xmm2, xmm0 // average columns (32 to 16 pixels)
psrlw xmm0, 8
movdqa xmm3, xmm1
psrlw xmm1, 8
pand xmm2, xmm5
pand xmm3, xmm5
pavgw xmm0, xmm2
pavgw xmm1, xmm3
packuswb xmm0, xmm1
sub ecx, 16
movdqa [edx], xmm0
lea edx, [edx + 16]
jg wloop
ret
}
}
// Blends 32x2 rectangle to 16x1. // Blends 32x2 rectangle to 16x1.
// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned. // Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
__declspec(naked) __declspec(align(16)) __declspec(naked) __declspec(align(16))
...@@ -281,6 +307,44 @@ static void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ...@@ -281,6 +307,44 @@ static void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
} }
} }
// Blends 32x1 rectangle to 16x1.
// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
__declspec(naked) __declspec(align(16))
void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t,
uint8* dst_ptr, int dst_width) {
__asm {
mov eax, [esp + 4] // src_ptr
// src_stride
mov edx, [esp + 12] // dst_ptr
mov ecx, [esp + 16] // dst_width
pcmpeqb xmm5, xmm5 // generate mask 0x00ff00ff
psrlw xmm5, 8
align 16
wloop:
movdqu xmm0, [eax]
movdqu xmm1, [eax + 16]
lea eax, [eax + 32]
movdqa xmm2, xmm0 // average columns (32 to 16 pixels)
psrlw xmm0, 8
movdqa xmm3, xmm1
psrlw xmm1, 8
pand xmm2, xmm5
pand xmm3, xmm5
pavgw xmm0, xmm2
pavgw xmm1, xmm3
packuswb xmm0, xmm1
sub ecx, 16
movdqu [edx], xmm0
lea edx, [edx + 16]
jg wloop
ret
}
}
// Blends 32x2 rectangle to 16x1. // Blends 32x2 rectangle to 16x1.
// Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned. // Alignment requirement: src_ptr 16 byte aligned, dst_ptr 16 byte aligned.
__declspec(naked) __declspec(align(16)) __declspec(naked) __declspec(align(16))
...@@ -838,6 +902,40 @@ static void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -838,6 +902,40 @@ static void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
); );
} }
void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t,
uint8* dst_ptr, int dst_width) {
asm volatile (
"pcmpeqb %%xmm5,%%xmm5 \n"
"psrlw $0x8,%%xmm5 \n"
".p2align 4 \n"
"1: \n"
"movdqa (%0),%%xmm0 \n"
"movdqa 0x10(%0),%%xmm1 \n"
"lea 0x20(%0),%0 \n"
"movdqa %%xmm0,%%xmm2 \n"
"psrlw $0x8,%%xmm0 \n"
"movdqa %%xmm1,%%xmm3 \n"
"psrlw $0x8,%%xmm1 \n"
"pand %%xmm5,%%xmm2 \n"
"pand %%xmm5,%%xmm3 \n"
"pavgw %%xmm2,%%xmm0 \n"
"pavgw %%xmm3,%%xmm1 \n"
"packuswb %%xmm1,%%xmm0 \n"
"movdqa %%xmm0,(%1) \n"
"lea 0x10(%1),%1 \n"
"sub $0x10,%2 \n"
"jg 1b \n"
: "+r"(src_ptr), // %0
"+r"(dst_ptr), // %1
"+r"(dst_width) // %2
:
: "memory", "cc"
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm5"
#endif
);
}
void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) { uint8* dst_ptr, int dst_width) {
asm volatile ( asm volatile (
...@@ -903,6 +1001,40 @@ static void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr, ...@@ -903,6 +1001,40 @@ static void ScaleRowDown2_Unaligned_SSE2(const uint8* src_ptr,
); );
} }
static void ScaleRowDown2Linear_Unaligned_SSE2(const uint8* src_ptr, ptrdiff_t,
uint8* dst_ptr, int dst_width) {
asm volatile (
"pcmpeqb %%xmm5,%%xmm5 \n"
"psrlw $0x8,%%xmm5 \n"
".p2align 4 \n"
"1: \n"
"movdqu (%0),%%xmm0 \n"
"movdqu 0x10(%0),%%xmm1 \n"
"lea 0x20(%0),%0 \n"
"movdqa %%xmm0,%%xmm2 \n"
"psrlw $0x8,%%xmm0 \n"
"movdqa %%xmm1,%%xmm3 \n"
"psrlw $0x8,%%xmm1 \n"
"pand %%xmm5,%%xmm2 \n"
"pand %%xmm5,%%xmm3 \n"
"pavgw %%xmm2,%%xmm0 \n"
"pavgw %%xmm3,%%xmm1 \n"
"packuswb %%xmm1,%%xmm0 \n"
"movdqu %%xmm0,(%1) \n"
"lea 0x10(%1),%1 \n"
"sub $0x10,%2 \n"
"jg 1b \n"
: "+r"(src_ptr), // %0
"+r"(dst_ptr), // %1
"+r"(dst_width) // %2
:
: "memory", "cc"
#if defined(__SSE2__)
, "xmm0", "xmm1", "xmm5"
#endif
);
}
static void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr, static void ScaleRowDown2Box_Unaligned_SSE2(const uint8* src_ptr,
ptrdiff_t src_stride, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) { uint8* dst_ptr, int dst_width) {
...@@ -1447,6 +1579,21 @@ static void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t /* src_stride */, ...@@ -1447,6 +1579,21 @@ static void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t /* src_stride */,
} }
} }
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, void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst, int dst_width) { uint8* dst, int dst_width) {
const uint8* s = src_ptr; const uint8* s = src_ptr;
...@@ -1685,7 +1832,9 @@ static void ScalePlaneDown2(int /* src_width */, int /* src_height */, ...@@ -1685,7 +1832,9 @@ static void ScalePlaneDown2(int /* src_width */, int /* src_height */,
FilterMode filtering) { FilterMode filtering) {
void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride, void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
uint8* dst_ptr, int dst_width) = uint8* dst_ptr, int dst_width) =
filtering ? ScaleRowDown2Box_C : ScaleRowDown2_C; filtering == kFilterNone ? ScaleRowDown2_C :
(filtering == kFilterLinear ? ScaleRowDown2Linear_C :
ScaleRowDown2Box_C);
int row_stride = src_stride << 1; int row_stride = src_stride << 1;
if (!filtering) { if (!filtering) {
src_ptr += src_stride; // Point to odd rows. src_ptr += src_stride; // Point to odd rows.
...@@ -1698,12 +1847,15 @@ static void ScalePlaneDown2(int /* src_width */, int /* src_height */, ...@@ -1698,12 +1847,15 @@ static void ScalePlaneDown2(int /* src_width */, int /* src_height */,
} }
#elif defined(HAS_SCALEROWDOWN2_SSE2) #elif defined(HAS_SCALEROWDOWN2_SSE2)
if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) { if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
ScaleRowDown2 = filtering ? ScaleRowDown2Box_Unaligned_SSE2 : ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Unaligned_SSE2 :
ScaleRowDown2_Unaligned_SSE2; (filtering == kFilterLinear ? ScaleRowDown2Linear_Unaligned_SSE2 :
ScaleRowDown2Box_Unaligned_SSE2);
if (IS_ALIGNED(src_ptr, 16) && if (IS_ALIGNED(src_ptr, 16) &&
IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) && IS_ALIGNED(src_stride, 16) && IS_ALIGNED(row_stride, 16) &&
IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) { IS_ALIGNED(dst_ptr, 16) && IS_ALIGNED(dst_stride, 16)) {
ScaleRowDown2 = filtering ? ScaleRowDown2Box_SSE2 : ScaleRowDown2_SSE2; ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
(filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
ScaleRowDown2Box_SSE2);
} }
} }
#elif defined(HAS_SCALEROWDOWN2_MIPS_DSPR2) #elif defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
...@@ -1715,6 +1867,9 @@ static void ScalePlaneDown2(int /* src_width */, int /* src_height */, ...@@ -1715,6 +1867,9 @@ static void ScalePlaneDown2(int /* src_width */, int /* src_height */,
} }
#endif #endif
if (filtering == kFilterLinear) {
src_stride = 0;
}
// TODO(fbarchard): Loop through source height to allow odd height. // TODO(fbarchard): Loop through source height to allow odd height.
for (int y = 0; y < dst_height; ++y) { for (int y = 0; y < dst_height; ++y) {
ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
...@@ -1759,6 +1914,9 @@ static void ScalePlaneDown4(int /* src_width */, int /* src_height */, ...@@ -1759,6 +1914,9 @@ static void ScalePlaneDown4(int /* src_width */, int /* src_height */,
} }
#endif #endif
if (filtering == kFilterLinear) {
src_stride = 0;
}
for (int y = 0; y < dst_height; ++y) { for (int y = 0; y < dst_height; ++y) {
ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
src_ptr += row_stride; src_ptr += row_stride;
...@@ -1822,14 +1980,15 @@ static void ScalePlaneDown34(int /* src_width */, int /* src_height */, ...@@ -1822,14 +1980,15 @@ static void ScalePlaneDown34(int /* src_width */, int /* src_height */,
} }
#endif #endif
const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
for (int y = 0; y < dst_height - 2; y += 3) { for (int y = 0; y < dst_height - 2; y += 3) {
ScaleRowDown34_0(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride; src_ptr += src_stride;
dst_ptr += dst_stride; dst_ptr += dst_stride;
ScaleRowDown34_1(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride; src_ptr += src_stride;
dst_ptr += dst_stride; dst_ptr += dst_stride;
ScaleRowDown34_0(src_ptr + src_stride, -src_stride, ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
dst_ptr, dst_width); dst_ptr, dst_width);
src_ptr += src_stride * 2; src_ptr += src_stride * 2;
dst_ptr += dst_stride; dst_ptr += dst_stride;
...@@ -1837,7 +1996,7 @@ static void ScalePlaneDown34(int /* src_width */, int /* src_height */, ...@@ -1837,7 +1996,7 @@ static void ScalePlaneDown34(int /* src_width */, int /* src_height */,
// Remainder 1 or 2 rows with last row vertically unfiltered // Remainder 1 or 2 rows with last row vertically unfiltered
if ((dst_height % 3) == 2) { if ((dst_height % 3) == 2) {
ScaleRowDown34_0(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride; src_ptr += src_stride;
dst_ptr += dst_stride; dst_ptr += dst_stride;
ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width); ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
...@@ -1914,21 +2073,22 @@ static void ScalePlaneDown38(int /* src_width */, int /* src_height */, ...@@ -1914,21 +2073,22 @@ static void ScalePlaneDown38(int /* src_width */, int /* src_height */,
} }
#endif #endif
const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
for (int y = 0; y < dst_height - 2; y += 3) { for (int y = 0; y < dst_height - 2; y += 3) {
ScaleRowDown38_3(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride * 3; src_ptr += src_stride * 3;
dst_ptr += dst_stride; dst_ptr += dst_stride;
ScaleRowDown38_3(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride * 3; src_ptr += src_stride * 3;
dst_ptr += dst_stride; dst_ptr += dst_stride;
ScaleRowDown38_2(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride * 2; src_ptr += src_stride * 2;
dst_ptr += dst_stride; dst_ptr += dst_stride;
} }
// Remainder 1 or 2 rows with last row vertically unfiltered // Remainder 1 or 2 rows with last row vertically unfiltered
if ((dst_height % 3) == 2) { if ((dst_height % 3) == 2) {
ScaleRowDown38_3(src_ptr, src_stride, dst_ptr, dst_width); ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
src_ptr += src_stride * 3; src_ptr += src_stride * 3;
dst_ptr += dst_stride; dst_ptr += dst_stride;
ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width); ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
...@@ -2080,7 +2240,8 @@ SAFEBUFFERS ...@@ -2080,7 +2240,8 @@ SAFEBUFFERS
void ScalePlaneBilinear(int src_width, int src_height, void ScalePlaneBilinear(int src_width, int src_height,
int dst_width, int dst_height, int dst_width, int dst_height,
int src_stride, int dst_stride, int src_stride, int dst_stride,
const uint8* src_ptr, uint8* dst_ptr) { const uint8* src_ptr, uint8* dst_ptr,
FilterMode filtering) {
assert(dst_width > 0); assert(dst_width > 0);
assert(dst_height > 0); assert(dst_height > 0);
assert(Abs(src_width) <= kMaxStride); assert(Abs(src_width) <= kMaxStride);
...@@ -2164,10 +2325,14 @@ void ScalePlaneBilinear(int src_width, int src_height, ...@@ -2164,10 +2325,14 @@ void ScalePlaneBilinear(int src_width, int src_height,
y = max_y; y = max_y;
} }
int yi = y >> 16; int yi = y >> 16;
int yf = (y >> 8) & 255;
const uint8* src = src_ptr + yi * src_stride; const uint8* src = src_ptr + yi * src_stride;
InterpolateRow(row, src, src_stride, src_width, yf); if (filtering == kFilterLinear) {
ScaleFilterCols_C(dst_ptr, row, dst_width, x, dx); ScaleFilterCols_C(dst_ptr, src, dst_width, x, dx);
} else {
int yf = (y >> 8) & 255;
InterpolateRow(row, src, src_stride, src_width, yf);
ScaleFilterCols_C(dst_ptr, row, dst_width, x, dx);
}
dst_ptr += dst_stride; dst_ptr += dst_stride;
y += dy; y += dy;
} }
...@@ -2219,15 +2384,11 @@ static void ScalePlaneAnySize(int src_width, int src_height, ...@@ -2219,15 +2384,11 @@ static void ScalePlaneAnySize(int src_width, int src_height,
src_stride, dst_stride, src_ptr, dst_ptr); src_stride, dst_stride, src_ptr, dst_ptr);
} else { } else {
ScalePlaneBilinear(src_width, src_height, dst_width, dst_height, ScalePlaneBilinear(src_width, src_height, dst_width, dst_height,
src_stride, dst_stride, src_ptr, dst_ptr); src_stride, dst_stride, src_ptr, dst_ptr, filtering);
} }
} }
// Scale plane down, any size // Scale plane down, any size
//
// This is an optimized version for scaling down a plane to any size.
// The current implementation is ~10 times faster compared to the
// reference implementation for e.g. XGA->LowResPAL
static void ScalePlaneDown(int src_width, int src_height, static void ScalePlaneDown(int src_width, int src_height,
int dst_width, int dst_height, int dst_width, int dst_height,
...@@ -2237,10 +2398,11 @@ static void ScalePlaneDown(int src_width, int src_height, ...@@ -2237,10 +2398,11 @@ static void ScalePlaneDown(int src_width, int src_height,
if (!filtering || src_width > kMaxStride) { if (!filtering || src_width > kMaxStride) {
ScalePlaneSimple(src_width, src_height, dst_width, dst_height, ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
src_stride, dst_stride, src_ptr, dst_ptr); src_stride, dst_stride, src_ptr, dst_ptr);
} else if (filtering == kFilterBilinear || dst_height * 2 > src_height) { } else if (filtering == kFilterBilinear || filtering == kFilterLinear ||
dst_height * 2 > src_height) {
// between 1/2x and 1x use bilinear // between 1/2x and 1x use bilinear
ScalePlaneBilinear(src_width, src_height, dst_width, dst_height, ScalePlaneBilinear(src_width, src_height, dst_width, dst_height,
src_stride, dst_stride, src_ptr, dst_ptr); src_stride, dst_stride, src_ptr, dst_ptr, filtering);
} else { } else {
ScalePlaneBox(src_width, src_height, dst_width, dst_height, ScalePlaneBox(src_width, src_height, dst_width, dst_height,
src_stride, dst_stride, src_ptr, dst_ptr); src_stride, dst_stride, src_ptr, dst_ptr);
...@@ -2271,12 +2433,8 @@ void ScalePlane(const uint8* src, int src_stride, ...@@ -2271,12 +2433,8 @@ void ScalePlane(const uint8* src, int src_stride,
0, 0, dy, 1, filtering); 0, 0, dy, 1, filtering);
} else if (dst_width <= Abs(src_width) && dst_height <= src_height) { } else if (dst_width <= Abs(src_width) && dst_height <= src_height) {
// Scale down. // Scale down.
if (use_reference_impl_) { if (4 * dst_width == 3 * src_width &&
// For testing, allow the optimized versions to be disabled. 4 * dst_height == 3 * src_height) {
ScalePlaneDown(src_width, src_height, dst_width, dst_height,
src_stride, dst_stride, src, dst, filtering);
} else if (4 * dst_width == 3 * src_width &&
4 * dst_height == 3 * src_height) {
// optimized, 3/4 // optimized, 3/4
ScalePlaneDown34(src_width, src_height, dst_width, dst_height, ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
src_stride, dst_stride, src, dst, filtering); src_stride, dst_stride, src, dst, filtering);
......
source/scale_argb.cc
View file @ 788f7570
...@@ -74,6 +74,36 @@ static void ScaleARGBRowDown2_SSE2(const uint8* src_argb, ...@@ -74,6 +74,36 @@ static void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
} }
} }
// 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. // Blends 8x2 rectangle to 4x1.
// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned. // Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
__declspec(naked) __declspec(align(16)) __declspec(naked) __declspec(align(16))
...@@ -466,6 +496,35 @@ static void ScaleARGBRowDown2_SSE2(const uint8* src_argb, ...@@ -466,6 +496,35 @@ static void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
); );
} }
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, static void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
ptrdiff_t src_stride, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) { uint8* dst_argb, int dst_width) {
...@@ -822,6 +881,19 @@ static void ScaleARGBRowDown2_C(const uint8* src_argb, ...@@ -822,6 +881,19 @@ static void ScaleARGBRowDown2_C(const uint8* src_argb,
} }
} }
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, static void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) { uint8* dst_argb, int dst_width) {
for (int x = 0; x < dst_width; ++x) { for (int x = 0; x < dst_width; ++x) {
...@@ -930,13 +1002,16 @@ static void ScaleARGBDown2(int /* src_width */, int /* src_height */, ...@@ -930,13 +1002,16 @@ static void ScaleARGBDown2(int /* src_width */, int /* src_height */,
int row_stride = src_stride * (dy >> 16); int row_stride = src_stride * (dy >> 16);
void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride, void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
uint8* dst_argb, int dst_width) = uint8* dst_argb, int dst_width) =
filtering ? ScaleARGBRowDown2Box_C : ScaleARGBRowDown2_C; filtering == kFilterNone ? ScaleARGBRowDown2_C :
(filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
ScaleARGBRowDown2Box_C);
#if defined(HAS_SCALEARGBROWDOWN2_SSE2) #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) && if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 4) &&
IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) && IS_ALIGNED(src_argb, 16) && IS_ALIGNED(row_stride, 16) &&
IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) { IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
ScaleARGBRowDown2 = filtering ? ScaleARGBRowDown2Box_SSE2 : ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
ScaleARGBRowDown2_SSE2; (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
ScaleARGBRowDown2Box_SSE2);
} }
#elif defined(HAS_SCALEARGBROWDOWN2_NEON) #elif defined(HAS_SCALEARGBROWDOWN2_NEON)
if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) && if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8) &&
...@@ -946,7 +1021,9 @@ static void ScaleARGBDown2(int /* src_width */, int /* src_height */, ...@@ -946,7 +1021,9 @@ static void ScaleARGBDown2(int /* src_width */, int /* src_height */,
} }
#endif #endif
// TODO(fbarchard): Loop through source height to allow odd height. if (filtering == kFilterLinear) {
src_stride = 0;
}
for (int y = 0; y < dst_height; ++y) { for (int y = 0; y < dst_height; ++y) {
ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width); ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width);
src_argb += row_stride; src_argb += row_stride;
...@@ -985,6 +1062,9 @@ static void ScaleARGBDownEven(int src_width, int src_height, ...@@ -985,6 +1062,9 @@ static void ScaleARGBDownEven(int src_width, int src_height,
} }
#endif #endif
if (filtering == kFilterLinear) {
src_stride = 0;
}
for (int y = 0; y < dst_height; ++y) { for (int y = 0; y < dst_height; ++y) {
ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width); ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width);
src_argb += row_stride; src_argb += row_stride;
...@@ -998,7 +1078,8 @@ static void ScaleARGBBilinearDown(int src_height, ...@@ -998,7 +1078,8 @@ static void ScaleARGBBilinearDown(int src_height,
int dst_width, int dst_height, int dst_width, int dst_height,
int src_stride, int dst_stride, int src_stride, int dst_stride,
const uint8* src_argb, uint8* dst_argb, const uint8* src_argb, uint8* dst_argb,
int x, int dx, int y, int dy) { int x, int dx, int y, int dy,
FilterMode filtering) {
assert(src_height > 0); assert(src_height > 0);
assert(dst_width > 0); assert(dst_width > 0);
assert(dst_height > 0); assert(dst_height > 0);
...@@ -1076,10 +1157,14 @@ static void ScaleARGBBilinearDown(int src_height, ...@@ -1076,10 +1157,14 @@ static void ScaleARGBBilinearDown(int src_height,
y = max_y; y = max_y;
} }
int yi = y >> 16; int yi = y >> 16;
int yf = (y >> 8) & 255;
const uint8* src = src_argb + yi * src_stride; const uint8* src = src_argb + yi * src_stride;
InterpolateRow(row, src, src_stride, clip_src_width, yf); if (filtering == kFilterLinear) {
ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx); ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
} else {
int yf = (y >> 8) & 255;
InterpolateRow(row, src, src_stride, clip_src_width, yf);
ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx);
}
dst_argb += dst_stride; dst_argb += dst_stride;
y += dy; y += dy;
} }
...@@ -1091,7 +1176,8 @@ static void ScaleARGBBilinearUp(int src_width, int src_height, ...@@ -1091,7 +1176,8 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
int dst_width, int dst_height, int dst_width, int dst_height,
int src_stride, int dst_stride, int src_stride, int dst_stride,
const uint8* src_argb, uint8* dst_argb, const uint8* src_argb, uint8* dst_argb,
int x, int dx, int y, int dy) { int x, int dx, int y, int dy,
FilterMode filtering) {
assert(src_width > 0); assert(src_width > 0);
assert(src_height > 0); assert(src_height > 0);
assert(dst_width > 0); assert(dst_width > 0);
...@@ -1180,8 +1266,12 @@ static void ScaleARGBBilinearUp(int src_width, int src_height, ...@@ -1180,8 +1266,12 @@ static void ScaleARGBBilinearUp(int src_width, int src_height,
src += src_stride; src += src_stride;
} }
} }
int yf = (y >> 8) & 255; if (filtering == kFilterLinear) {
InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf); InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
} else {
int yf = (y >> 8) & 255;
InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
}
dst_argb += dst_stride; dst_argb += dst_stride;
y += dy; y += dy;
} }
...@@ -1200,7 +1290,8 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height, ...@@ -1200,7 +1290,8 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
const uint8* src_u, const uint8* src_u,
const uint8* src_v, const uint8* src_v,
uint8* dst_argb, uint8* dst_argb,
int x, int dx, int y, int dy) { int x, int dx, int y, int dy,
FilterMode filtering) {
assert(src_width > 0); assert(src_width > 0);
assert(src_height > 0); assert(src_height > 0);
assert(dst_width > 0); assert(dst_width > 0);
...@@ -1353,8 +1444,12 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height, ...@@ -1353,8 +1444,12 @@ static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
} }
} }
} }
int yf = (y >> 8) & 255; if (filtering == kFilterLinear) {
InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf); InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
} else {
int yf = (y >> 8) & 255;
InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
}
dst_argb += dst_stride_argb; dst_argb += dst_stride_argb;
y += dy; y += dy;
} }
...@@ -1424,14 +1519,14 @@ static void ScaleARGBAnySize(int src_width, int src_height, ...@@ -1424,14 +1519,14 @@ static void ScaleARGBAnySize(int src_width, int src_height,
ScaleARGBBilinearUp(src_width, src_height, ScaleARGBBilinearUp(src_width, src_height,
clip_width, clip_height, clip_width, clip_height,
src_stride, dst_stride, src_argb, dst_argb, src_stride, dst_stride, src_argb, dst_argb,
x, dx, y, dy); x, dx, y, dy, filtering);
return; return;
} }
if (filtering && src_width * 4 < kMaxStride) { if (filtering && src_width * 4 < kMaxStride) {
ScaleARGBBilinearDown(src_height, ScaleARGBBilinearDown(src_height,
clip_width, clip_height, clip_width, clip_height,
src_stride, dst_stride, src_argb, dst_argb, src_stride, dst_stride, src_argb, dst_argb,
x, dx, y, dy); x, dx, y, dy, filtering);
return; return;
} }
ScaleARGBSimple(src_width, src_height, clip_width, clip_height, ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
......
unit_test/scale_argb_test.cc
View file @ 788f7570
...@@ -213,18 +213,20 @@ static int ARGBClipTestFilter(int src_width, int src_height, ...@@ -213,18 +213,20 @@ static int ARGBClipTestFilter(int src_width, int src_height,
// Test a scale factor with 2 filters. Expect unfiltered to be exact, but // Test a scale factor with 2 filters. Expect unfiltered to be exact, but
// filtering is different fixed point implementations for SSSE3, Neon and C. // filtering is different fixed point implementations for SSSE3, Neon and C.
#define TEST_FACTOR(name, hfactor, vfactor) \ #define TEST_FACTOR(name, hfactor, vfactor) \
TEST_FACTOR1(name, None, hfactor, vfactor, 2) \
TEST_FACTOR1(name, Linear, hfactor, vfactor, 2) \
TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2)
// TODO(fbarchard): ScaleDownBy1 should be lossless, but Box has error of 2. // TODO(fbarchard): ScaleDownBy1 should be lossless, but Box has error of 2.
// TEST_FACTOR(1, 1 / 1, 1 / 1) TEST_FACTOR(1, 1 / 1, 1 / 1)
TEST_FACTOR(2, 1 / 2, 1 / 2) TEST_FACTOR(2, 1 / 2, 1 / 2)
TEST_FACTOR(4, 1 / 4, 1 / 4) TEST_FACTOR(4, 1 / 4, 1 / 4)
// TEST_FACTOR(8, 1 / 8, 1 / 8) TEST_FACTOR(8, 1 / 8, 1 / 8)
// TEST_FACTOR(16, 1 / 16, 1 / 16) TEST_FACTOR(16, 1 / 16, 1 / 16)
// TEST_FACTOR(2by3, 2 / 3, 2 / 3) TEST_FACTOR(2by3, 2 / 3, 2 / 3)
TEST_FACTOR(3by4, 3 / 4, 3 / 4) TEST_FACTOR(3by4, 3 / 4, 3 / 4)
// TEST_FACTOR(3by8, 3 / 8, 3 / 8) TEST_FACTOR(3by8, 3 / 8, 3 / 8)
// TEST_FACTOR(Vertical2by3, 1, 2 / 3) TEST_FACTOR(Vertical2by3, 1, 2 / 3)
#undef TEST_FACTOR1 #undef TEST_FACTOR1
#undef TEST_FACTOR #undef TEST_FACTOR
...@@ -257,14 +259,15 @@ TEST_FACTOR(3by4, 3 / 4, 3 / 4) ...@@ -257,14 +259,15 @@ TEST_FACTOR(3by4, 3 / 4, 3 / 4)
// Test scale to a specified size with all 3 filters. // Test scale to a specified size with all 3 filters.
#define TEST_SCALETO(name, width, height) \ #define TEST_SCALETO(name, width, height) \
TEST_SCALETO1(name, width, height, None, 0) \ TEST_SCALETO1(name, width, height, None, 0) \
TEST_SCALETO1(name, width, height, Linear, 2) \
TEST_SCALETO1(name, width, height, Bilinear, 2) TEST_SCALETO1(name, width, height, Bilinear, 2)
TEST_SCALETO(ARGBScale, 640, 360) TEST_SCALETO(ARGBScale, 640, 360)
TEST_SCALETO(DISABLED_ARGBScale, 853, 480) TEST_SCALETO(ARGBScale, 853, 480)
TEST_SCALETO(DISABLED_ARGBScale, 1280, 720) TEST_SCALETO(ARGBScale, 1280, 720)
TEST_SCALETO(DISABLED_ARGBScale, 1280, 800) TEST_SCALETO(ARGBScale, 1280, 800)
TEST_SCALETO(DISABLED_ARGBScale, 1366, 768) TEST_SCALETO(ARGBScale, 1366, 768)
TEST_SCALETO(DISABLED_ARGBScale, 1920, 1080) TEST_SCALETO(ARGBScale, 1920, 1080)
#undef TEST_SCALETO1 #undef TEST_SCALETO1
#undef TEST_SCALETO #undef TEST_SCALETO
......
unit_test/scale_test.cc
View file @ 788f7570
...@@ -141,23 +141,24 @@ static int TestFilter(int src_width, int src_height, ...@@ -141,23 +141,24 @@ static int TestFilter(int src_width, int src_height,
EXPECT_LE(diff, max_diff); \ EXPECT_LE(diff, max_diff); \
} }
// Test a scale factor with all 3 filters. Expect unfiltered to be exact, but // Test a scale factor with all 4 filters. Expect unfiltered to be exact, but
// filtering is different fixed point implementations for SSSE3, Neon and C. // filtering is different fixed point implementations for SSSE3, Neon and C.
#define TEST_FACTOR(name, hfactor, vfactor) \ #define TEST_FACTOR(name, hfactor, vfactor) \
TEST_FACTOR1(name, None, hfactor, vfactor, 0) \ TEST_FACTOR1(name, None, hfactor, vfactor, 0) \
TEST_FACTOR1(name, Linear, hfactor, vfactor, 2) \
TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) \ TEST_FACTOR1(name, Bilinear, hfactor, vfactor, 2) \
TEST_FACTOR1(name, Box, hfactor, vfactor, 2) \ TEST_FACTOR1(name, Box, hfactor, vfactor, 2) \
// TODO(fbarchard): ScaleDownBy1 should be lossless, but Box has error of 2. // TODO(fbarchard): ScaleDownBy1 should be lossless, but Box has error of 2.
// TEST_FACTOR(1, 1 / 1, 1 / 1) TEST_FACTOR(1, 1 / 1, 1 / 1)
TEST_FACTOR(2, 1 / 2, 1 / 2) TEST_FACTOR(2, 1 / 2, 1 / 2)
TEST_FACTOR(4, 1 / 4, 1 / 4) TEST_FACTOR(4, 1 / 4, 1 / 4)
// TEST_FACTOR(8, 1 / 8, 1 / 8) TEST_FACTOR(8, 1 / 8, 1 / 8)
// TEST_FACTOR(16, 1 / 16, 1 / 16) TEST_FACTOR(16, 1 / 16, 1 / 16)
// TEST_FACTOR(2by3, 2 / 3, 2 / 3) TEST_FACTOR(2by3, 2 / 3, 2 / 3)
TEST_FACTOR(3by4, 3 / 4, 3 / 4) TEST_FACTOR(3by4, 3 / 4, 3 / 4)
// TEST_FACTOR(3by8, 3 / 8, 3 / 8) TEST_FACTOR(3by8, 3 / 8, 3 / 8)
// TEST_FACTOR(Vertical2by3, 1, 2 / 3) TEST_FACTOR(Vertical2by3, 1, 2 / 3)
#undef TEST_FACTOR1 #undef TEST_FACTOR1
#undef TEST_FACTOR #undef TEST_FACTOR
...@@ -175,17 +176,19 @@ TEST_FACTOR(3by4, 3 / 4, 3 / 4) ...@@ -175,17 +176,19 @@ TEST_FACTOR(3by4, 3 / 4, 3 / 4)
EXPECT_LE(diff, max_diff); \ EXPECT_LE(diff, max_diff); \
} }
// Test scale to a specified size with all 3 filters. // Test scale to a specified size with all 4 filters.
#define TEST_SCALETO(name, width, height) \ #define TEST_SCALETO(name, width, height) \
TEST_SCALETO1(name, width, height, None, 0) \ TEST_SCALETO1(name, width, height, None, 0) \
TEST_SCALETO1(name, width, height, Linear, 0) \
TEST_SCALETO1(name, width, height, Bilinear, 2) \
TEST_SCALETO1(name, width, height, Box, 2) TEST_SCALETO1(name, width, height, Box, 2)
TEST_SCALETO(Scale, 640, 360) TEST_SCALETO(Scale, 640, 360)
TEST_SCALETO(DISABLED_Scale, 853, 480) TEST_SCALETO(Scale, 853, 480)
TEST_SCALETO(DISABLED_Scale, 1280, 720) TEST_SCALETO(Scale, 1280, 720)
TEST_SCALETO(DISABLED_Scale, 1280, 800) TEST_SCALETO(Scale, 1280, 800)
TEST_SCALETO(DISABLED_Scale, 1366, 768) TEST_SCALETO(Scale, 1366, 768)
TEST_SCALETO(DISABLED_Scale, 1920, 1080) TEST_SCALETO(Scale, 1920, 1080)
#undef TEST_SCALETO1 #undef TEST_SCALETO1
#undef TEST_SCALETO #undef TEST_SCALETO
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment