mirror of
https://gitlab.winehq.org/wine/wine-gecko.git
synced 2024-09-13 09:24:08 -07:00
67 lines
1.9 KiB
Diff
67 lines
1.9 KiB
Diff
|
diff --git a/gfx/ycbcr/yuv_row_linux.cpp b/gfx/ycbcr/yuv_row_linux.cpp
|
||
|
|
--- a/gfx/ycbcr/yuv_row_linux.cpp
|
||
|
|
+++ b/gfx/ycbcr/yuv_row_linux.cpp
|
||
|
|
@@ -250,18 +250,18 @@ MMX_ALIGNED(int16 kCoefficientsRgbY[768]
|
||
|
|
|
||
|
|
// AMD64 ABI uses register paremters.
|
||
|
|
void FastConvertYUVToRGB32Row(const uint8* y_buf, // rdi
|
||
|
|
const uint8* u_buf, // rsi
|
||
|
|
const uint8* v_buf, // rdx
|
||
|
|
uint8* rgb_buf, // rcx
|
||
|
|
int width) { // r8
|
||
|
|
asm(
|
||
|
|
- "jmp convertend\n"
|
||
|
|
-"convertloop:"
|
||
|
|
+ "jmp Lconvertend\n"
|
||
|
|
+"Lconvertloop:"
|
||
|
|
"movzb (%1),%%r10\n"
|
||
|
|
"add $0x1,%1\n"
|
||
|
|
"movzb (%2),%%r11\n"
|
||
|
|
"add $0x1,%2\n"
|
||
|
|
"movq 2048(%5,%%r10,8),%%xmm0\n"
|
||
|
|
"movzb (%0),%%r10\n"
|
||
|
|
"movq 4096(%5,%%r11,8),%%xmm1\n"
|
||
|
|
"movzb 0x1(%0),%%r11\n"
|
||
|
|
@@ -271,36 +271,36 @@ void FastConvertYUVToRGB32Row(const uint
|
||
|
|
"movq (%5,%%r11,8),%%xmm3\n"
|
||
|
|
"paddsw %%xmm0,%%xmm2\n"
|
||
|
|
"paddsw %%xmm0,%%xmm3\n"
|
||
|
|
"shufps $0x44,%%xmm3,%%xmm2\n"
|
||
|
|
"psraw $0x6,%%xmm2\n"
|
||
|
|
"packuswb %%xmm2,%%xmm2\n"
|
||
|
|
"movq %%xmm2,0x0(%3)\n"
|
||
|
|
"add $0x8,%3\n"
|
||
|
|
-"convertend:"
|
||
|
|
+"Lconvertend:"
|
||
|
|
"sub $0x2,%4\n"
|
||
|
|
- "jns convertloop\n"
|
||
|
|
+ "jns Lconvertloop\n"
|
||
|
|
|
||
|
|
-"convertnext:"
|
||
|
|
+"Lconvertnext:"
|
||
|
|
"add $0x1,%4\n"
|
||
|
|
- "js convertdone\n"
|
||
|
|
+ "js Lconvertdone\n"
|
||
|
|
|
||
|
|
"movzb (%1),%%r10\n"
|
||
|
|
"movq 2048(%5,%%r10,8),%%xmm0\n"
|
||
|
|
"movzb (%2),%%r10\n"
|
||
|
|
"movq 4096(%5,%%r10,8),%%xmm1\n"
|
||
|
|
"paddsw %%xmm1,%%xmm0\n"
|
||
|
|
"movzb (%0),%%r10\n"
|
||
|
|
"movq (%5,%%r10,8),%%xmm1\n"
|
||
|
|
"paddsw %%xmm0,%%xmm1\n"
|
||
|
|
"psraw $0x6,%%xmm1\n"
|
||
|
|
"packuswb %%xmm1,%%xmm1\n"
|
||
|
|
"movd %%xmm1,0x0(%3)\n"
|
||
|
|
-"convertdone:"
|
||
|
|
+"Lconvertdone:"
|
||
|
|
:
|
||
|
|
: "r"(y_buf), // %0
|
||
|
|
"r"(u_buf), // %1
|
||
|
|
"r"(v_buf), // %2
|
||
|
|
"r"(rgb_buf), // %3
|
||
|
|
"r"(width), // %4
|
||
|
|
"r" (kCoefficientsRgbY) // %5
|
||
|
|
: "memory", "r10", "r11", "xmm0", "xmm1", "xmm2", "xmm3"
|