--- a/gfx/ycbcr/yuv_row_linux.cpp
+++ b/gfx/ycbcr/yuv_row_linux.cpp
@@ -6,18 +6,18 @@
#define DCHECK(a)
// TODO(fbarchard): Move this to yuv_row_posix.cc to share with Mac.
// TODO(fbarchard): Do 64 bit version.
extern "C" {
-#if defined(ARCH_CPU_ARM_FAMILY)
-// ARM implementation uses C fallback
+#ifndef ARCH_CPU_X86_FAMILY
+// non-x86 implementation uses C fallback
void FastConvertYUVToRGB32Row(const uint8* y_buf,
const uint8* u_buf,
const uint8* v_buf,
uint8* rgb_buf,
int width) {
FastConvertYUVToRGB32Row_C(y_buf, u_buf, v_buf, rgb_buf, width, 1);
}
@@ -241,18 +241,82 @@ MMX_ALIGNED(int16 kCoefficientsRgbY[768]
RGBV(0xE8), RGBV(0xE9), RGBV(0xEA), RGBV(0xEB),
RGBV(0xEC), RGBV(0xED), RGBV(0xEE), RGBV(0xEF),
RGBV(0xF0), RGBV(0xF1), RGBV(0xF2), RGBV(0xF3),
RGBV(0xF4), RGBV(0xF5), RGBV(0xF6), RGBV(0xF7),
RGBV(0xF8), RGBV(0xF9), RGBV(0xFA), RGBV(0xFB),
RGBV(0xFC), RGBV(0xFD), RGBV(0xFE), RGBV(0xFF),
};
+#ifdef __SUNPRO_CC
+#pragma align 16 (kCoefficientsRgbY)
+#endif
+
#if defined(ARCH_CPU_X86_64)
+#ifdef __SUNPRO_CC
+// AMD64 ABI uses register parameters.
+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:"
+ "movzbq (%1),%%r10\n"
+ "add $0x1,%1\n"
+ "movzbq (%2),%%r11\n"
+ "add $0x1,%2\n"
+ "movq 2048(%5,%%r10,8),%%xmm0\n"
+ "movzbq (%0),%%r10\n"
+ "movq 4096(%5,%%r11,8),%%xmm1\n"
+ "movzbq 0x1(%0),%%r11\n"
+ "paddsw %%xmm1,%%xmm0\n"
+ "movq (%5,%%r10,8),%%xmm2\n"
+ "add $0x2,%0\n"
+ "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:"
+ "sub $0x2,%4\n"
+ "jns convertloop\n"
+
+"convertnext:"
+ "add $0x1,%4\n"
+ "js convertdone\n"
+
+ "movzbq (%1),%%r10\n"
+ "movq 2048(%5,%%r10,8),%%xmm0\n"
+ "movzbq (%2),%%r10\n"
+ "movq 4096(%5,%%r10,8),%%xmm1\n"
+ "paddsw %%xmm1,%%xmm0\n"
+ "movzbq (%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:"
+ :
+ : "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"
+);
+}
+#else // __SUNPRO_CC
// 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 1f\n"
@@ -301,19 +365,79 @@ void FastConvertYUVToRGB32Row(const uint
"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"
);
}
+#endif // __SUNPRO_CC
-#else
+#else // ARCH_CPU_X86_64
+
+#ifdef __SUNPRO_CC
+void FastConvertYUVToRGB32Row(const uint8* y_buf,
+ const uint8* u_buf,
+ const uint8* v_buf,
+ uint8* rgb_buf,
+ int width) {
+ asm(
+ "pusha\n"
+ "mov %eax,%ebp\n"
+ "jmp convertend\n"
+"convertloop:"
+ "movzbl (%edi),%eax\n"
+ "add $0x1,%edi\n"
+ "movzbl (%esi),%ebx\n"
+ "add $0x1,%esi\n"
+ "movq kCoefficientsRgbY+2048(,%eax,8),%mm0\n"
+ "movzbl (%edx),%eax\n"
+ "paddsw kCoefficientsRgbY+4096(,%ebx,8),%mm0\n"
+ "movzbl 0x1(%edx),%ebx\n"
+ "movq kCoefficientsRgbY(,%eax,8),%mm1\n"
+ "add $0x2,%edx\n"
+ "movq kCoefficientsRgbY(,%ebx,8),%mm2\n"
+ "paddsw %mm0,%mm1\n"
+ "paddsw %mm0,%mm2\n"
+ "psraw $0x6,%mm1\n"
+ "psraw $0x6,%mm2\n"
+ "packuswb %mm2,%mm1\n"
+ "movntq %mm1,0x0(%ebp)\n"
+ "add $0x8,%ebp\n"
+"convertend:"
+ "sub $0x2,%ecx\n"
+ "jns convertloop\n"
+
+ "and $0x1,%ecx\n"
+ "je convertdone\n"
+
+ "movzbl (%edi),%eax\n"
+ "movq kCoefficientsRgbY+2048(,%eax,8),%mm0\n"
+ "movzbl (%esi),%eax\n"
+ "paddsw kCoefficientsRgbY+4096(,%eax,8),%mm0\n"
+ "movzbl (%edx),%eax\n"
+ "movq kCoefficientsRgbY(,%eax,8),%mm1\n"
+ "paddsw %mm0,%mm1\n"
+ "psraw $0x6,%mm1\n"
+ "packuswb %mm1,%mm1\n"
+ "movd %mm1,0x0(%ebp)\n"
+"convertdone:"
+ "popa\n"
+ :
+ : "d"(y_buf), // %edx
+ "D"(u_buf), // %edi
+ "S"(v_buf), // %esi
+ "a"(rgb_buf), // %eax
+ "c"(width) // %ecx
+ : "memory"
+);
+}
+#else // __SUNPRO_CC
void FastConvertYUVToRGB32Row(const uint8* y_buf,
const uint8* u_buf,
const uint8* v_buf,
uint8* rgb_buf,
int width);
// It's necessary to specify the correct section for the following code,
// otherwise it will be placed in whatever the current section is as this unit
@@ -369,12 +493,13 @@ void FastConvertYUVToRGB32Row(const uint
"packuswb %mm1,%mm1\n"
"movd %mm1,0x0(%ebp)\n"
"2:"
"popa\n"
"ret\n"
".previous\n"
);
-#endif
-#endif // ARCH_CPU_ARM_FAMILY
+#endif // __SUNPRO_CC
+#endif // ARCH_CPU_X86_64
+#endif // !ARCH_CPU_X86_FAMILY
} // extern "C"
