Change the x86 idct functions to do reconstruction at the same time

Change-Id: I896fe6f9664e6849c7cee2cc6bb4e045eb42540f

3 files changed, 82 insertions, 25 deletions

diff --git a/vp8/decoder/x86/dequantize_mmx.asm b/vp8/decoder/x86/dequantize_mmx.asm
index 2dcf548f6..7ad9289cc 100644
--- a/vp8/decoder/x86/dequantize_mmx.asm
+++ b/vp8/decoder/x86/dequantize_mmx.asm
@@ -50,12 +50,12 @@ sym(vp8_dequantize_b_impl_mmx):
     ret
 
 
-;void dequant_idct_mmx(short *input, short *dq, short *output, int pitch)
-global sym(vp8_dequant_idct_mmx)
-sym(vp8_dequant_idct_mmx):
+;void dequant_idct_add_mmx(short *input, short *dq, unsigned char *pred, unsigned char *dest, int pitch, int stride)
+global sym(vp8_dequant_idct_add_mmx)
+sym(vp8_dequant_idct_add_mmx):
     push        rbp
     mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 4
+    SHADOW_ARGS_TO_STACK 6
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -77,7 +77,8 @@ sym(vp8_dequant_idct_mmx):
         movq        mm3,    [rax+24]
         pmullw      mm3,    [rdx+24]
 
-        mov         rdx,    arg(2) ;output
+        mov         rdx,    arg(3) ;dest
+        mov         rsi,    arg(2) ;pred
         pxor        mm7,    mm7
 
 
@@ -88,7 +89,8 @@ sym(vp8_dequant_idct_mmx):
         movq        [rax+24],mm7
 
 
-        movsxd      rax,            dword ptr arg(3) ;pitch
+        movsxd      rax,            dword ptr arg(4) ;pitch
+        movsxd      rdi,            dword ptr arg(5) ;stride
 
         psubw       mm0,            mm2             ; b1= 0-2
         paddw       mm2,            mm2             ;
@@ -207,13 +209,34 @@ sym(vp8_dequant_idct_mmx):
         punpckldq   mm2,            mm4             ; 32 22 12 02
         punpckhdq   mm5,            mm4             ; 33 23 13 03
 
-        movq        [rdx],          mm0
+        pxor        mm7,            mm7
 
-        movq        [rdx+rax],      mm1
-        movq        [rdx+rax*2],    mm2
+        movd        mm4,            [rsi]
+        punpcklbw   mm4,            mm7
+        paddsw      mm0,            mm4
+        packuswb    mm0,            mm7
+        movd        [rdx],          mm0
 
-        add         rdx,            rax
-        movq        [rdx+rax*2],    mm5
+        movd        mm4,            [rsi+rax]
+        punpcklbw   mm4,            mm7
+        paddsw      mm1,            mm4
+        packuswb    mm1,            mm7
+        movd        [rdx+rdi],      mm1
+
+        movd        mm4,            [rsi+2*rax]
+        punpcklbw   mm4,            mm7
+        paddsw      mm2,            mm4
+        packuswb    mm2,            mm7
+        movd        [rdx+rdi*2],    mm2
+
+        add         rdx,            rdi
+        add         rsi,            rax
+
+        movd        mm4,            [rsi+2*rax]
+        punpcklbw   mm4,            mm7
+        paddsw      mm5,            mm4
+        packuswb    mm5,            mm7
+        movd        [rdx+rdi*2],    mm5
 
     ; begin epilog
     pop rdi
@@ -224,12 +247,12 @@ sym(vp8_dequant_idct_mmx):
     ret
 
 
-;void dequant_dc_idct_mmx(short *input, short *dq, short *output, int pitch, int Dc)
-global sym(vp8_dequant_dc_idct_mmx)
-sym(vp8_dequant_dc_idct_mmx):
+;void dequant_dc_idct_add_mmx(short *input, short *dq, unsigned char *pred, unsigned char *dest, int pitch, int stride, int Dc)
+global sym(vp8_dequant_dc_idct_add_mmx)
+sym(vp8_dequant_dc_idct_add_mmx):
     push        rbp
     mov         rbp, rsp
-    SHADOW_ARGS_TO_STACK 5
+    SHADOW_ARGS_TO_STACK 7
     GET_GOT     rbx
     push        rsi
     push        rdi
@@ -238,7 +261,7 @@ sym(vp8_dequant_dc_idct_mmx):
         mov         rax,    arg(0) ;input
         mov         rdx,    arg(1) ;dq
 
-        movsxd      rcx,    dword ptr arg(4) ;Dc
+        movsxd      rcx,    dword ptr arg(6) ;Dc
 
         movq        mm0,    [rax   ]
         pmullw      mm0,    [rdx]
@@ -252,7 +275,8 @@ sym(vp8_dequant_dc_idct_mmx):
         movq        mm3,    [rax+24]
         pmullw      mm3,    [rdx+24]
 
-        mov         rdx,    arg(2) ;output
+        mov         rdx,    arg(3) ;dest
+        mov         rsi,    arg(2) ;pred
         pxor        mm7,    mm7
 
 
@@ -262,8 +286,10 @@ sym(vp8_dequant_dc_idct_mmx):
         movq        [rax+16],mm7
         movq        [rax+24],mm7
 
+
         pinsrw      mm0,    rcx,  0
-        movsxd      rax,            dword ptr arg(3) ;pitch
+        movsxd      rax,            dword ptr arg(4) ;pitch
+        movsxd      rdi,            dword ptr arg(5) ;stride
 
         psubw       mm0,            mm2             ; b1= 0-2
         paddw       mm2,            mm2             ;
@@ -382,13 +408,34 @@ sym(vp8_dequant_dc_idct_mmx):
         punpckldq   mm2,            mm4             ; 32 22 12 02
         punpckhdq   mm5,            mm4             ; 33 23 13 03
 
-        movq        [rdx],          mm0
-
-        movq        [rdx+rax],      mm1
-        movq        [rdx+rax*2],    mm2
-
-        add         rdx,            rax
-        movq        [rdx+rax*2],    mm5
+        pxor        mm7,            mm7
+
+        movd        mm4,            [rsi]
+        punpcklbw   mm4,            mm7
+        paddsw      mm0,            mm4
+        packuswb    mm0,            mm7
+        movd        [rdx],          mm0
+
+        movd        mm4,            [rsi+rax]
+        punpcklbw   mm4,            mm7
+        paddsw      mm1,            mm4
+        packuswb    mm1,            mm7
+        movd        [rdx+rdi],      mm1
+
+        movd        mm4,            [rsi+2*rax]
+        punpcklbw   mm4,            mm7
+        paddsw      mm2,            mm4
+        packuswb    mm2,            mm7
+        movd        [rdx+rdi*2],    mm2
+
+        add         rdx,            rdi
+        add         rsi,            rax
+
+        movd        mm4,            [rsi+2*rax]
+        punpcklbw   mm4,            mm7
+        paddsw      mm5,            mm4
+        packuswb    mm5,            mm7
+        movd        [rdx+rdi*2],    mm5
 
     ; begin epilog
     pop rdi
diff --git a/vp8/decoder/x86/dequantize_x86.h b/vp8/decoder/x86/dequantize_x86.h
index 563d58460..32e777994 100644
--- a/vp8/decoder/x86/dequantize_x86.h
+++ b/vp8/decoder/x86/dequantize_x86.h
@@ -21,12 +21,20 @@
  */
 #if HAVE_MMX
 extern prototype_dequant_block(vp8_dequantize_b_mmx);
+extern prototype_dequant_idct_add(vp8_dequant_idct_add_mmx);
+extern prototype_dequant_idct_dc_add(vp8_dequant_dc_idct_add_mmx);
 
 
 #if !CONFIG_RUNTIME_CPU_DETECT
 #undef  vp8_dequant_block
 #define vp8_dequant_block vp8_dequantize_b_mmx
 
+#undef  vp8_dequant_idct_add
+#define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
+
+#undef  vp8_dequant_idct_dc
+#define vp8_dequant_idct_add_dc vp8_dequant_dc_idct_add_mmx
+
 #endif
 #endif
 
diff --git a/vp8/decoder/x86/x86_dsystemdependent.c b/vp8/decoder/x86/x86_dsystemdependent.c
index bcd2acc75..d7bed0873 100644
--- a/vp8/decoder/x86/x86_dsystemdependent.c
+++ b/vp8/decoder/x86/x86_dsystemdependent.c
@@ -43,6 +43,8 @@ void vp8_arch_x86_decode_init(VP8D_COMP *pbi)
     if (flags & HAS_MMX)
     {
         pbi->dequant.block   = vp8_dequantize_b_mmx;
+        pbi->dequant.idct_add    = vp8_dequant_idct_add_mmx;
+        pbi->dequant.idct_dc_add = vp8_dequant_dc_idct_add_mmx;
     }
 
 #endif