Use lea to remove 2 instructions from the inner loop of YUV Scale.

Use MMX2 to remove emms.
unittest has code to time yuv scale and convert 100 times so it takes about 1 second... useful for catching performance issues.  But the code is disabled.
BUG=23263
TEST=media_unittest should pass and run much faster: 28% faster than 32 bit version.

Review URL: http://codereview.chromium.org/273047

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@29019 0039d316-1c4b-4281-b951-d872f2087c98

1 files changed, 38 insertions, 40 deletions

diff --git a/media/base/yuv_row_linux.cc b/media/base/yuv_row_linux.cc
index 5825960..b99c7bf 100644
--- a/media/base/yuv_row_linux.cc
+++ b/media/base/yuv_row_linux.cc
@@ -263,19 +263,19 @@ void FastConvertYUVToRGB32Row(const uint8* y_buf,  // rdi
   "add    $0x1,%rsi\n"
   "movzb  (%rdx),%r11\n"
   "add    $0x1,%rdx\n"
-  "movq   kCoefficientsRgbU(,%r10,8),%mm0\n"
+  "movq   kCoefficientsRgbU(,%r10,8),%xmm0\n"
   "movzb  (%rdi),%r10\n"
-  "paddsw kCoefficientsRgbV(,%r11,8),%mm0\n"
+  "paddsw kCoefficientsRgbV(,%r11,8),%xmm0\n"
   "movzb  0x1(%rdi),%r11\n"
-  "movq   kCoefficientsRgbY(,%r10,8),%mm1\n"
+  "movq   kCoefficientsRgbY(,%r10,8),%xmm1\n"
   "add    $0x2,%rdi\n"
-  "movq   kCoefficientsRgbY(,%r11,8),%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(%rcx)\n"
+  "movq   kCoefficientsRgbY(,%r11,8),%xmm2\n"
+  "paddsw %xmm0,%xmm1\n"
+  "paddsw %xmm0,%xmm2\n"
+  "psraw  $0x6,%xmm1\n"
+  "psraw  $0x6,%xmm2\n"
+  "packuswb %xmm2,%xmm1\n"
+  "movq   %xmm1,0x0(%rcx)\n"
   "add    $0x8,%rcx\n"
 "convertend:"
   "sub    $0x2,%r8\n"
@@ -285,15 +285,15 @@ void FastConvertYUVToRGB32Row(const uint8* y_buf,  // rdi
   "je     convertdone\n"
 
   "movzb  (%rsi),%r10\n"
-  "movq   kCoefficientsRgbU(,%r10,8),%mm0\n"
+  "movq   kCoefficientsRgbU(,%r10,8),%xmm0\n"
   "movzb  (%rdx),%r10\n"
-  "paddsw kCoefficientsRgbV(,%r10,8),%mm0\n"
+  "paddsw kCoefficientsRgbV(,%r10,8),%xmm0\n"
   "movzb  (%rdi),%r10\n"
-  "movq   kCoefficientsRgbY(,%r10,8),%mm1\n"
-  "paddsw %mm0,%mm1\n"
-  "psraw  $0x6,%mm1\n"
-  "packuswb %mm1,%mm1\n"
-  "movd   %mm1,0x0(%rcx)\n"
+  "movq   kCoefficientsRgbY(,%r10,8),%xmm1\n"
+  "paddsw %xmm0,%xmm1\n"
+  "psraw  $0x6,%xmm1\n"
+  "packuswb %xmm1,%xmm1\n"
+  "movd   %xmm1,0x0(%rcx)\n"
 "convertdone:"
   "ret\n"
 );
@@ -316,25 +316,23 @@ void ScaleYUVToRGB32Row(const uint8* y_buf,  // rdi
   "mov    %r11,%r10\n"
   "sar    $0x5,%r10\n"
   "movzb  (%rsi,%r10,1),%rax\n"
-  "movq   kCoefficientsRgbU(,%rax,8),%mm0\n"
+  "movq   kCoefficientsRgbU(,%rax,8),%xmm0\n"
   "movzb  (%rdx,%r10,1),%rax\n"
-  "paddsw kCoefficientsRgbV(,%rax,8),%mm0\n"
-  "mov    %r11,%r10\n"
-  "add    %r9,%r11\n"
-  "sar    $0x4,%r10\n"
-  "movzb  (%rdi,%r10,1),%rax\n"
-  "movq   kCoefficientsRgbY(,%rax,8),%mm1\n"
-  "mov    %r11,%r10\n"
-  "add    %r9,%r11\n"
+  "paddsw kCoefficientsRgbV(,%rax,8),%xmm0\n"
+  "lea    (%r11,%r9),%r10\n"
+  "sar    $0x4,%r11\n"
+  "movzb  (%rdi,%r11,1),%rax\n"
+  "movq   kCoefficientsRgbY(,%rax,8),%xmm1\n"
+  "lea    (%r10,%r9),%r11\n"
   "sar    $0x4,%r10\n"
   "movzb  (%rdi,%r10,1),%rax\n"
-  "movq   kCoefficientsRgbY(,%rax,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(%rcx)\n"
+  "movq   kCoefficientsRgbY(,%rax,8),%xmm2\n"
+  "paddsw %xmm0,%xmm1\n"
+  "paddsw %xmm0,%xmm2\n"
+  "psraw  $0x6,%xmm1\n"
+  "psraw  $0x6,%xmm2\n"
+  "packuswb %xmm2,%xmm1\n"
+  "movq   %xmm1,0x0(%rcx)\n"
   "add    $0x8,%rcx\n"
 "scaleend:"
   "sub    $0x2,%r8\n"
@@ -346,16 +344,16 @@ void ScaleYUVToRGB32Row(const uint8* y_buf,  // rdi
   "mov    %r11,%r10\n"
   "sar    $0x5,%r10\n"
   "movzb  (%rsi,%r10,1),%rax\n"
-  "movq   kCoefficientsRgbU(,%rax,8),%mm0\n"
+  "movq   kCoefficientsRgbU(,%rax,8),%xmm0\n"
   "movzb  (%rdx,%r10,1),%rax\n"
-  "paddsw kCoefficientsRgbV(,%rax,8),%mm0\n"
+  "paddsw kCoefficientsRgbV(,%rax,8),%xmm0\n"
   "sar    $0x4,%r11\n"
   "movzb  (%rdi,%r11,1),%rax\n"
-  "movq   kCoefficientsRgbY(,%rax,8),%mm1\n"
-  "paddsw %mm0,%mm1\n"
-  "psraw  $0x6,%mm1\n"
-  "packuswb %mm1,%mm1\n"
-  "movd   %mm1,0x0(%rcx)\n"
+  "movq   kCoefficientsRgbY(,%rax,8),%xmm1\n"
+  "paddsw %xmm0,%xmm1\n"
+  "psraw  $0x6,%xmm1\n"
+  "packuswb %xmm1,%xmm1\n"
+  "movd   %xmm1,0x0(%rcx)\n"
 
 "scaledone:"
   "ret\n"