Skip to content

L
libyuv
Commit 14f657b4 authored by fbarchard@google.com's avatar fbarchard@google.com
Browse files
Options

SSE2 scale for ARGB and posix with specializations 

BUG=177
TESTED=try bots
Review URL: https://webrtc-codereview.appspot.com/1115008

git-svn-id: http://libyuv.googlecode.com/svn/trunk@585 16f28f9a-4ce2-e073-06de-1de4eb20be90
parent b3446fc5
Hide whitespace changes
Inline Side-by-side
Showing with 166 additions and 62 deletions
README.chromium
View file @ 14f657b4
Name: libyuv Name: libyuv
URL: http://code.google.com/p/libyuv/ URL: http://code.google.com/p/libyuv/
Version: 584 Version: 585
License: BSD License: BSD
License File: LICENSE License File: LICENSE
......
include/libyuv/row.h
View file @ 14f657b4
...@@ -169,7 +169,8 @@ extern "C" { ...@@ -169,7 +169,8 @@ extern "C" {
#endif #endif
// The following are available on Neon platforms // The following are available on Neon platforms
#if !defined(LIBYUV_DISABLE_NEON) && (defined(__ARM_NEON__) || defined(LIBYUV_NEON)) #if !defined(LIBYUV_DISABLE_NEON) && \
(defined(__ARM_NEON__) || defined(LIBYUV_NEON))
#define HAS_ABGRTOARGBROW_NEON #define HAS_ABGRTOARGBROW_NEON
#define HAS_ABGRTOUVROW_NEON #define HAS_ABGRTOUVROW_NEON
#define HAS_ABGRTOYROW_NEON #define HAS_ABGRTOYROW_NEON
......
include/libyuv/version.h
View file @ 14f657b4
...@@ -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 584 #define LIBYUV_VERSION 585
#endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT #endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT
source/scale.cc
View file @ 14f657b4
...@@ -23,9 +23,6 @@ namespace libyuv { ...@@ -23,9 +23,6 @@ namespace libyuv {
extern "C" { extern "C" {
#endif #endif
// Bilinear SSE2 is disabled.
#define SSE2_DISABLED 1
// 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
...@@ -1954,20 +1951,29 @@ static void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride, ...@@ -1954,20 +1951,29 @@ static void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
static void ScaleFilterRows_SSE2(uint8* dst_ptr, static void ScaleFilterRows_SSE2(uint8* dst_ptr,
const uint8* src_ptr, ptrdiff_t src_stride, const uint8* src_ptr, ptrdiff_t src_stride,
int dst_width, int source_y_fraction) { int dst_width, int source_y_fraction) {
asm volatile (
asm volatile ( asm volatile (
"sub %1,%0 \n" "sub %1,%0 \n"
"shr %3 \n"
"cmp $0x0,%3 \n" "cmp $0x0,%3 \n"
"je 2f \n" "je 100f \n"
"cmp $0x80,%3 \n" "cmp $0x20,%3 \n"
"je 3f \n" "je 75f \n"
"cmp $0x40,%3 \n"
"je 50f \n"
"cmp $0x60,%3 \n"
"je 25f \n"
"movd %3,%%xmm0 \n"
"neg %3 \n"
"add $0x80,%3 \n"
"movd %3,%%xmm5 \n" "movd %3,%%xmm5 \n"
"punpcklbw %%xmm5,%%xmm5 \n" "punpcklbw %%xmm0,%%xmm5 \n"
"psrlw $0x1,%%xmm5 \n"
"punpcklwd %%xmm5,%%xmm5 \n" "punpcklwd %%xmm5,%%xmm5 \n"
"punpckldq %%xmm5,%%xmm5 \n" "pshufd $0x0,%%xmm5,%%xmm5 \n"
"punpcklqdq %%xmm5,%%xmm5 \n"
"pxor %%xmm4,%%xmm4 \n" "pxor %%xmm4,%%xmm4 \n"
// General purpose row blend.
".p2align 4 \n" ".p2align 4 \n"
"1: \n" "1: \n"
"movdqa (%1),%%xmm0 \n" "movdqa (%1),%%xmm0 \n"
...@@ -1991,25 +1997,56 @@ static void ScaleFilterRows_SSE2(uint8* dst_ptr, ...@@ -1991,25 +1997,56 @@ static void ScaleFilterRows_SSE2(uint8* dst_ptr,
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 1b \n" "jg 1b \n"
"jmp 4f \n" "jmp 99f \n"
// Blend 25 / 75.
".p2align 4 \n" ".p2align 4 \n"
"2: \n" "25: \n"
"movdqa (%1),%%xmm0 \n" "movdqa (%1),%%xmm0 \n"
"movdqa (%1,%4,1),%%xmm1 \n"
"pavgb %%xmm1,%%xmm0 \n"
"pavgb %%xmm1,%%xmm0 \n"
"sub $0x10,%2 \n" "sub $0x10,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 2b \n" "jg 25b \n"
"jmp 4f \n" "jmp 99f \n"
// Blend 50 / 50.
".p2align 4 \n" ".p2align 4 \n"
"3: \n" "50: \n"
"movdqa (%1),%%xmm0 \n" "movdqa (%1),%%xmm0 \n"
"pavgb (%1,%4,1),%%xmm0 \n" "movdqa (%1,%4,1),%%xmm1 \n"
"pavgb %%xmm1,%%xmm0 \n"
"sub $0x10,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n"
"jg 50b \n"
"jmp 99f \n"
// Blend 75 / 25.
".p2align 4 \n"
"75: \n"
"movdqa (%1),%%xmm1 \n"
"movdqa (%1,%4,1),%%xmm0 \n"
"pavgb %%xmm1,%%xmm0 \n"
"pavgb %%xmm1,%%xmm0 \n"
"sub $0x10,%2 \n" "sub $0x10,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 3b \n" "jg 75b \n"
"jmp 99f \n"
// Blend 100 / 0 - Copy row unchanged.
".p2align 4 \n" ".p2align 4 \n"
"4: \n" "100: \n"
"movdqa (%1),%%xmm0 \n"
"sub $0x10,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n"
"jg 100b \n"
"99: \n"
"punpckhbw %%xmm0,%%xmm0 \n" "punpckhbw %%xmm0,%%xmm0 \n"
"pshufhw $0xff,%%xmm0,%%xmm0 \n" "pshufhw $0xff,%%xmm0,%%xmm0 \n"
"punpckhqdq %%xmm0,%%xmm0 \n" "punpckhqdq %%xmm0,%%xmm0 \n"
......
source/scale_argb.cc
View file @ 14f657b4
...@@ -200,8 +200,7 @@ static void ScaleARGBRowDownEvenInt_SSE2(const uint8* src_argb, ...@@ -200,8 +200,7 @@ static void ScaleARGBRowDownEvenInt_SSE2(const uint8* src_argb,
} }
// Bilinear row filtering combines 4x2 -> 4x1. SSE2 version. // Bilinear row filtering combines 4x2 -> 4x1. SSE2 version.
#ifndef SSE2_DISABLED #define HAS_SCALEARGBFILTERROWS_SSE2
#define HAS_SCALEARGBFILTERROWS_SSE2_DISABLED
__declspec(naked) __declspec(align(16)) __declspec(naked) __declspec(align(16))
void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
ptrdiff_t src_stride, int dst_width, ptrdiff_t src_stride, int dst_width,
...@@ -215,19 +214,24 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -215,19 +214,24 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
mov ecx, [esp + 8 + 16] // dst_width mov ecx, [esp + 8 + 16] // dst_width
mov eax, [esp + 8 + 20] // source_y_fraction (0..255) mov eax, [esp + 8 + 20] // source_y_fraction (0..255)
sub edi, esi sub edi, esi
// Dispatch to specialized filters if applicable.
cmp eax, 0 cmp eax, 0
je xloop1 je xloop100 // 0 / 256. Blend 100 / 0.
cmp eax, 64
je xloop75 // 64 / 256 is 0.25. Blend 75 / 25.
cmp eax, 128 cmp eax, 128
je xloop2 je xloop50 // 128 / 256 is 0.50. Blend 50 / 50.
cmp eax, 192
je xloop25 // 192 / 256 is 0.75. Blend 25 / 75.
movd xmm5, eax // xmm5 = y fraction movd xmm5, eax // xmm5 = y fraction
punpcklbw xmm5, xmm5 punpcklbw xmm5, xmm5
psrlw xmm5, 1
punpcklwd xmm5, xmm5 punpcklwd xmm5, xmm5
pshufd xmm5, xmm5, 0 punpckldq xmm5, xmm5
punpcklqdq xmm5, xmm5
pxor xmm4, xmm4 pxor xmm4, xmm4
// f * row1 + (1 - frac) row0
// frac * (row1 - row0) + row0
align 16 align 16
xloop: xloop:
movdqa xmm0, [esi] // row0 movdqa xmm0, [esi] // row0
...@@ -240,6 +244,8 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -240,6 +244,8 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
punpckhbw xmm1, xmm4 punpckhbw xmm1, xmm4
psubw xmm2, xmm0 // row1 - row0 psubw xmm2, xmm0 // row1 - row0
psubw xmm3, xmm1 psubw xmm3, xmm1
paddw xmm2, xmm2 // 9 bits * 15 bits = 8.16
paddw xmm3, xmm3
pmulhw xmm2, xmm5 // scale diff pmulhw xmm2, xmm5 // scale diff
pmulhw xmm3, xmm5 pmulhw xmm3, xmm5
paddw xmm0, xmm2 // sum rows paddw xmm0, xmm2 // sum rows
...@@ -249,44 +255,63 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -249,44 +255,63 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
movdqa [esi + edi], xmm0 movdqa [esi + edi], xmm0
lea esi, [esi + 16] lea esi, [esi + 16]
jg xloop jg xloop
jmp xloop99
shufps xmm0, xmm0, 0xff // Blend 25 / 75.
movdqa [esi + edi], xmm0 // duplicate last pixel for filtering align 16
pop edi xloop25:
pop esi movdqa xmm0, [esi]
ret movdqa xmm1, [esi + edx]
pavgb xmm0, xmm1
pavgb xmm0, xmm1
sub ecx, 4
movdqa [esi + edi], xmm0
lea esi, [esi + 16]
jg xloop25
jmp xloop99
// Blend 50 / 50.
align 16 align 16
xloop1: xloop50:
movdqa xmm0, [esi] movdqa xmm0, [esi]
movdqa xmm1, [esi + edx]
pavgb xmm0, xmm1
sub ecx, 4 sub ecx, 4
movdqa [esi + edi], xmm0 movdqa [esi + edi], xmm0
lea esi, [esi + 16] lea esi, [esi + 16]
jg xloop1 jg xloop50
jmp xloop99
shufps xmm0, xmm0, 0xff // Blend 75 / 25.
align 16
xloop75:
movdqa xmm1, [esi]
movdqa xmm0, [esi + edx]
pavgb xmm0, xmm1
pavgb xmm0, xmm1
sub ecx, 4
movdqa [esi + edi], xmm0 movdqa [esi + edi], xmm0
pop edi lea esi, [esi + 16]
pop esi jg xloop75
ret jmp xloop99
// Blend 100 / 0 - Copy row unchanged.
align 16 align 16
xloop2: xloop100:
movdqa xmm0, [esi] movdqa xmm0, [esi]
pavgb xmm0, [esi + edx]
sub ecx, 4 sub ecx, 4
movdqa [esi + edi], xmm0 movdqa [esi + edi], xmm0
lea esi, [esi + 16] lea esi, [esi + 16]
jg xloop2 jg xloop100
xloop99:
shufps xmm0, xmm0, 0xff shufps xmm0, xmm0, 0xff
movdqa [esi + edi], xmm0 movdqa [esi + edi], xmm0 // duplicate last pixel for filtering
pop edi pop edi
pop esi pop esi
ret ret
} }
} }
#endif // SSE2_DISABLED
// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version. // Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version.
#define HAS_SCALEARGBFILTERROWS_SSSE3 #define HAS_SCALEARGBFILTERROWS_SSSE3
...@@ -544,23 +569,33 @@ static void ScaleARGBRowDownEvenInt_SSE2(const uint8* src_argb, ...@@ -544,23 +569,33 @@ static void ScaleARGBRowDownEvenInt_SSE2(const uint8* src_argb,
); );
} }
#ifndef SSE2_DISABLED
// Bilinear row filtering combines 4x2 -> 4x1. SSE2 version // Bilinear row filtering combines 4x2 -> 4x1. SSE2 version
#define HAS_SCALEARGBFILTERROWS_SSE2_DISABLED #define HAS_SCALEARGBFILTERROWS_SSE2
void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
ptrdiff_t src_stride, int dst_width, ptrdiff_t src_stride, int dst_width,
int source_y_fraction) { int source_y_fraction) {
asm volatile ( asm volatile (
"sub %1,%0 \n" "sub %1,%0 \n"
"shr %3 \n"
"cmp $0x0,%3 \n" "cmp $0x0,%3 \n"
"je 2f \n" "je 100f \n"
"cmp $0x80,%3 \n" "cmp $0x20,%3 \n"
"je 3f \n" "je 75f \n"
"cmp $0x40,%3 \n"
"je 50f \n"
"cmp $0x60,%3 \n"
"je 25f \n"
"movd %3,%%xmm0 \n"
"neg %3 \n"
"add $0x80,%3 \n"
"movd %3,%%xmm5 \n" "movd %3,%%xmm5 \n"
"punpcklbw %%xmm5,%%xmm5 \n" "punpcklbw %%xmm0,%%xmm5 \n"
"punpcklwd %%xmm5,%%xmm5 \n" "punpcklwd %%xmm5,%%xmm5 \n"
"pshufd $0x0,%%xmm5,%%xmm5 \n" "pshufd $0x0,%%xmm5,%%xmm5 \n"
"pxor %%xmm4,%%xmm4 \n" "pxor %%xmm4,%%xmm4 \n"
// General purpose row blend.
".p2align 4 \n" ".p2align 4 \n"
"1: \n" "1: \n"
"movdqa (%1),%%xmm0 \n" "movdqa (%1),%%xmm0 \n"
...@@ -573,6 +608,8 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -573,6 +608,8 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
"punpckhbw %%xmm4,%%xmm1 \n" "punpckhbw %%xmm4,%%xmm1 \n"
"psubw %%xmm0,%%xmm2 \n" "psubw %%xmm0,%%xmm2 \n"
"psubw %%xmm1,%%xmm3 \n" "psubw %%xmm1,%%xmm3 \n"
"paddw %%xmm2,%%xmm2 \n"
"paddw %%xmm3,%%xmm3 \n"
"pmulhw %%xmm5,%%xmm2 \n" "pmulhw %%xmm5,%%xmm2 \n"
"pmulhw %%xmm5,%%xmm3 \n" "pmulhw %%xmm5,%%xmm3 \n"
"paddw %%xmm2,%%xmm0 \n" "paddw %%xmm2,%%xmm0 \n"
...@@ -582,31 +619,61 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -582,31 +619,61 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 1b \n" "jg 1b \n"
"jmp 4f \n" "jmp 99f \n"
// Blend 25 / 75.
".p2align 4 \n" ".p2align 4 \n"
"2: \n" "25: \n"
"movdqa (%1),%%xmm0 \n" "movdqa (%1),%%xmm0 \n"
"movdqa (%1,%4,1),%%xmm1 \n"
"pavgb %%xmm1,%%xmm0 \n"
"pavgb %%xmm1,%%xmm0 \n"
"sub $0x4,%2 \n" "sub $0x4,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 2b \n" "jg 25b \n"
"jmp 4f \n" "jmp 99f \n"
// Blend 50 / 50.
".p2align 4 \n" ".p2align 4 \n"
"3: \n" "50: \n"
"movdqa (%1),%%xmm0 \n" "movdqa (%1),%%xmm0 \n"
"pavgb (%1,%4,1),%%xmm0 \n" "movdqa (%1,%4,1),%%xmm1 \n"
"pavgb %%xmm1,%%xmm0 \n"
"sub $0x4,%2 \n" "sub $0x4,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 50b \n"
"jmp 99f \n"
// Blend 75 / 25.
".p2align 4 \n"
"75: \n"
"movdqa (%1),%%xmm1 \n"
"movdqa (%1,%4,1),%%xmm0 \n"
"pavgb %%xmm1,%%xmm0 \n"
"pavgb %%xmm1,%%xmm0 \n"
"sub $0x4,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n" "lea 0x10(%1),%1 \n"
"jg 3b \n" "jg 75b \n"
"jmp 99f \n"
// Blend 100 / 0 - Copy row unchanged.
".p2align 4 \n" ".p2align 4 \n"
"4: \n" "100: \n"
"movdqa (%1),%%xmm0 \n"
"sub $0x4,%2 \n"
"movdqa %%xmm0,(%1,%0,1) \n"
"lea 0x10(%1),%1 \n"
"jg 100b \n"
"99: \n"
"shufps $0xff,%%xmm0,%%xmm0 \n" "shufps $0xff,%%xmm0,%%xmm0 \n"
"movdqa %%xmm0,(%1,%0,1) \n" "movdqa %%xmm0,(%1,%0,1) \n"
: "+r"(dst_argb), // %0 : "+r"(dst_argb), // %0
"+r"(src_argb), // %1 "+r"(src_argb), // %1
"+r"(dst_width), // %2 "+r"(dst_width), // %2
"+r"(source_y_fraction) // %3 "+r"(source_y_fraction) // %3
: "r"(static_cast<intptr_t>(src_stride)) // %4 : "r"(static_cast<intptr_t>(src_stride)) // %4
: "memory", "cc" : "memory", "cc"
...@@ -615,7 +682,6 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb, ...@@ -615,7 +682,6 @@ void ScaleARGBFilterRows_SSE2(uint8* dst_argb, const uint8* src_argb,
#endif #endif
); );
} }
#endif // SSE2_DISABLED
// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version // Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
#define HAS_SCALEARGBFILTERROWS_SSSE3 #define HAS_SCALEARGBFILTERROWS_SSSE3
......
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