Adjust the reference window based on MVP prediction to optimize VME param on Haswell

Signed-off-by: Zhao Yakui <yakui.zhao@intel.com>

5 files changed, 57 insertions, 10 deletions

diff --git a/src/shaders/vme/inter_bframe_haswell.asm b/src/shaders/vme/inter_bframe_haswell.asm
index 9ab35d3..e02748e 100644
--- a/src/shaders/vme/inter_bframe_haswell.asm
+++ b/src/shaders/vme/inter_bframe_haswell.asm
@@ -78,6 +78,7 @@ send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen
 
 mov  (8) vme_m1.0<1>:ud		0:ud		{align1};
 mov  (8) mb_mvp_ref.0<1>:ud	0:ud		{align1};
+mov  (8) mb_ref_win.0<1>:ud	0:ud		{align1};
 and.z.f0.0 (1)		null:uw	mb_hwdep<0,1,0>:uw		0x04:uw   {align1};
 (f0.0) jmpi (1) __mb_hwdep_end;
 
@@ -462,6 +463,9 @@ jmpi	(1)	word_imedian;
 mov	(1)	mb_mvp_ref.6<1>:w		RET_ARG<0,1,0>:w	{align1};
 
 __mb_hwdep_end:
+asr	(4)	mb_ref_win.0<1>:w	mb_mvp_ref.0<4,4,1>:w	2:w	{align1};
+add	(4)	mb_ref_win.8<1>:w	mb_ref_win.0<4,4,1>:w	3:w	{align1};
+and	(4)	mb_ref_win.16<1>:uw	mb_ref_win.8<4,4,1>:uw	0xFFFC:uw {align1};
 /* m2, get the MV/Mb cost passed from constant buffer when
 spawning thread by MEDIA_OBJECT */       
 mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
@@ -580,7 +584,16 @@ mov  (1) vme_m0.0<1>:W		-8:W			{align1};
 mov  (1) vme_m0.2<1>:W		-8:W			{align1};
 
 mov  (1) vme_m0.4<1>:UD		vme_m0.0<0,1,0>:UD	{align1};
+and.z.f0.0 (1)		null:uw	input_mb_intra_ub<0,1,0>:ub	INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
+(f0.0)	add 	(1)	vme_m0.0<1>:w	vme_m0.0<0,1,0>:w	4:w	{align1};
+(f0.0)	add 	(1)	vme_m0.4<1>:w	vme_m0.4<0,1,0>:w	4:w	{align1};
+and.z.f0.0 (1)		null:uw	input_mb_intra_ub<0,1,0>:ub	INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
+(f0.0)	add 	(1)	vme_m0.2<1>:w	vme_m0.2<0,1,0>:w	4:w	{align1};
+(f0.0)	add 	(1)	vme_m0.6<1>:w	vme_m0.6<0,1,0>:w	4:w	{align1};
 
+add  (2) vme_m0.0<1>:w		vme_m0.0<2,2,1>:w	mb_ref_win.16<2,2,1>:w	{align1};
+add  (2) vme_m0.4<1>:w		vme_m0.4<2,2,1>:w	mb_ref_win.20<2,2,1>:w	{align1};
+ 
 mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
 
 mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
diff --git a/src/shaders/vme/inter_bframe_haswell.g75b b/src/shaders/vme/inter_bframe_haswell.g75b
index 6327a32..e0ef98e 100644
--- a/src/shaders/vme/inter_bframe_haswell.g75b
+++ b/src/shaders/vme/inter_bframe_haswell.g75b
@@ -35,6 +35,7 @@
    { 0x04600031, 0x26201cb1, 0x00000800, 0x02190006 },
    { 0x00600001, 0x24600061, 0x00000000, 0x00000000 },
    { 0x00600001, 0x2ac00061, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2a800061, 0x00000000, 0x00000000 },
    { 0x01000005, 0x20002d28, 0x000000a6, 0x00040004 },
    { 0x00010020, 0x34001c00, 0x00001400, 0x00000cb0 },
    { 0x00600001, 0x2ae00061, 0x00000000, 0x00000000 },
@@ -63,7 +64,7 @@
    { 0x00000001, 0x2fa40021, 0x00000b80, 0x00000000 },
    { 0x00000001, 0x2fa80061, 0x00000000, 0x00000001 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x000012f0 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x000013a0 },
    { 0x00000001, 0x2aa00129, 0x00000fe4, 0x00000000 },
    { 0x01000010, 0x20002d28, 0x00000aa0, 0x00000000 },
    { 0x00010001, 0x2af001e9, 0x00000000, 0x00010001 },
@@ -99,7 +100,7 @@
    { 0x00000001, 0x2fa40021, 0x00000b80, 0x00000000 },
    { 0x00000001, 0x2fa80061, 0x00000000, 0x00000002 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x000010b0 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x00001160 },
    { 0x00000001, 0x2aa00129, 0x00000fe4, 0x00000000 },
    { 0x01000010, 0x20002d28, 0x00000aa0, 0x00000000 },
    { 0x00010001, 0x2b1001e9, 0x00000000, 0x00010001 },
@@ -134,7 +135,7 @@
    { 0x00000001, 0x2fa40021, 0x00000b80, 0x00000000 },
    { 0x00000001, 0x2fa80061, 0x00000000, 0x00000002 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x00000e80 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x00000f30 },
    { 0x00000001, 0x2aa00129, 0x00000fe4, 0x00000000 },
    { 0x01000010, 0x20002d28, 0x00000aa0, 0x00000000 },
    { 0x00010001, 0x2b3001e9, 0x00000000, 0x00010001 },
@@ -166,7 +167,7 @@
    { 0x00000001, 0x2fa40021, 0x00000b80, 0x00000000 },
    { 0x00000001, 0x2fa80061, 0x00000000, 0x00000003 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x00000c80 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x00000d30 },
    { 0x00000001, 0x2aa00129, 0x00000fe4, 0x00000000 },
    { 0x00000001, 0x2b3201ed, 0x00000000, 0x00010001 },
    { 0x01000010, 0x20002d28, 0x00000aa0, 0x00000000 },
@@ -207,13 +208,13 @@
    { 0x00000001, 0x2fa401ad, 0x00000b04, 0x00000000 },
    { 0x00000001, 0x2fa801ad, 0x00000b24, 0x00000000 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x000008f0 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x000009a0 },
    { 0x00000001, 0x2ac001ad, 0x00000fe4, 0x00000000 },
    { 0x00000001, 0x2fa001ad, 0x00000ae6, 0x00000000 },
    { 0x00000001, 0x2fa401ad, 0x00000b06, 0x00000000 },
    { 0x00000001, 0x2fa801ad, 0x00000b26, 0x00000000 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x00000890 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x00000940 },
    { 0x00000001, 0x2ac201ad, 0x00000fe4, 0x00000000 },
    { 0x00600001, 0x24000061, 0x00000000, 0x00000000 },
    { 0x01000010, 0x20003da4, 0x00200af6, 0x00000000 },
@@ -232,14 +233,17 @@
    { 0x00000001, 0x2fa401ad, 0x00000b08, 0x00000000 },
    { 0x00000001, 0x2fa801ad, 0x00000b28, 0x00000000 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x00000760 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x00000810 },
    { 0x00000001, 0x2ac401ad, 0x00000fe4, 0x00000000 },
    { 0x00000001, 0x2fa001ad, 0x00000aea, 0x00000000 },
    { 0x00000001, 0x2fa401ad, 0x00000b0a, 0x00000000 },
    { 0x00000001, 0x2fa801ad, 0x00000b2a, 0x00000000 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x00000700 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x000007b0 },
    { 0x00000001, 0x2ac601ad, 0x00000fe4, 0x00000000 },
+   { 0x0040000c, 0x2a803dad, 0x00690ac0, 0x00020002 },
+   { 0x00400040, 0x2a883dad, 0x00690a80, 0x00030003 },
+   { 0x00400005, 0x2a902d29, 0x00690a88, 0xfffcfffc },
    { 0x00600001, 0x25600021, 0x008d0020, 0x00000000 },
    { 0x00600001, 0x28400021, 0x008d0560, 0x00000000 },
    { 0x00600001, 0x28600061, 0x00000000, 0x00000000 },
@@ -284,6 +288,14 @@
    { 0x00000001, 0x244001ed, 0x00000000, 0xfff8fff8 },
    { 0x00000001, 0x244201ed, 0x00000000, 0xfff8fff8 },
    { 0x00000001, 0x24440021, 0x00000440, 0x00000000 },
+   { 0x01000005, 0x20002e28, 0x000000a5, 0x00600060 },
+   { 0x00010040, 0x24403dad, 0x00000440, 0x00040004 },
+   { 0x00010040, 0x24443dad, 0x00000444, 0x00040004 },
+   { 0x01000005, 0x20002e28, 0x000000a5, 0x00100010 },
+   { 0x00010040, 0x24423dad, 0x00000442, 0x00040004 },
+   { 0x00010040, 0x24463dad, 0x00000446, 0x00040004 },
+   { 0x00200040, 0x244035ad, 0x00450440, 0x00450a90 },
+   { 0x00200040, 0x244435ad, 0x00450444, 0x00450a94 },
    { 0x00600001, 0x28000021, 0x008d0440, 0x00000000 },
    { 0x00000001, 0x24600061, 0x00000000, 0x00000002 },
    { 0x00000001, 0x24640231, 0x0000009c, 0x00000000 },
diff --git a/src/shaders/vme/inter_frame_haswell.asm b/src/shaders/vme/inter_frame_haswell.asm
index 36b394a..fa9a0a0 100644
--- a/src/shaders/vme/inter_frame_haswell.asm
+++ b/src/shaders/vme/inter_frame_haswell.asm
@@ -77,6 +77,7 @@ mov  (8) msg_reg0.0<1>:UD       read1_header.0<8,8,1>:UD {align1};
 send (8) msg_ind CHROMA_COL<1>:UB null read(BIND_IDX_CBCR, 0, 0, 4) mlen 1 rlen 1 {align1};
 
 mov  (8) mb_mvp_ref.0<1>:ud	0:ud		{align1};
+mov  (8) mb_ref_win.0<1>:ud	0:ud		{align1};
 and.z.f0.0 (1)		null:uw	mb_hwdep<0,1,0>:uw		0x04:uw   {align1};
 (f0.0) jmpi (1) __mb_hwdep_end;
 /* read back the data for MB A */
@@ -364,6 +365,9 @@ jmpi	(1)	word_imedian;
 mov	(1)	mb_mvp_ref.2<1>:w		RET_ARG<0,1,0>:w	{align1};
 
 __mb_hwdep_end:
+asr	(2)	mb_ref_win.0<1>:w	mb_mvp_ref.0<2,2,1>:w	2:w	{align1};
+add	(2)	mb_ref_win.8<1>:w	mb_ref_win.0<2,2,1>:w	3:w	{align1};
+and	(2)	mb_ref_win.16<1>:uw	mb_ref_win.8<2,2,1>:uw	0xFFFC:uw {align1};
 /* m2, get the MV/Mb cost passed from constant buffer when
 spawning thread by MEDIA_OBJECT */       
 mov (8) vme_m2<1>:UD            r1.0<8,8,1>:UD {align1};
@@ -481,6 +485,13 @@ mov  (1) vme_m0.2<1>:W		-12:W			{align1};
 
 mov  (1) vme_m0.4<1>:UD		vme_m0.0<0,1,0>:UD	{align1};
 
+and.z.f0.0 (1)		null:uw	input_mb_intra_ub<0,1,0>:ub	INTRA_PRED_AVAIL_FLAG_AE:uw   {align1};
+(f0.0)	add 	(1)	vme_m0.0<1>:w	vme_m0.0<0,1,0>:w	12:w	{align1};
+and.z.f0.0 (1)		null:uw	input_mb_intra_ub<0,1,0>:ub	INTRA_PRED_AVAIL_FLAG_B:uw   {align1};
+(f0.0)	add 	(1)	vme_m0.2<1>:w	vme_m0.2<0,1,0>:w	8:w	{align1};
+ 
+add  (2) vme_m0.0<1>:w		vme_m0.0<2,2,1>:w	mb_ref_win.16<2,2,1>:w	{align1};
+add  (2) vme_m0.4<1>:w		vme_m0.4<2,2,1>:w	mb_ref_win.16<2,2,1>:w	{align1};
 mov  (8) vme_msg_0.0<1>:UD      vme_m0.0<8,8,1>:UD {align1};
 
 mov  (1) vme_m1.0<1>:UD         ADAPTIVE_SEARCH_ENABLE:ud {align1} ;
diff --git a/src/shaders/vme/inter_frame_haswell.g75b b/src/shaders/vme/inter_frame_haswell.g75b
index 1ef526c..2ef1826 100644
--- a/src/shaders/vme/inter_frame_haswell.g75b
+++ b/src/shaders/vme/inter_frame_haswell.g75b
@@ -34,6 +34,7 @@
    { 0x00600001, 0x28000021, 0x008d0420, 0x00000000 },
    { 0x04600031, 0x26201cb1, 0x00000800, 0x02190006 },
    { 0x00600001, 0x2ac00061, 0x00000000, 0x00000000 },
+   { 0x00600001, 0x2a800061, 0x00000000, 0x00000000 },
    { 0x01000005, 0x20002d28, 0x000000a6, 0x00040004 },
    { 0x00010020, 0x34001c00, 0x00001400, 0x00000710 },
    { 0x00600001, 0x2ae00061, 0x00000000, 0x00000000 },
@@ -141,14 +142,17 @@
    { 0x00000001, 0x2fa401ad, 0x00000b04, 0x00000000 },
    { 0x00000001, 0x2fa801ad, 0x00000b24, 0x00000000 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x000007b0 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x00000840 },
    { 0x00000001, 0x2ac001ad, 0x00000fe4, 0x00000000 },
    { 0x00000001, 0x2fa001ad, 0x00000ae6, 0x00000000 },
    { 0x00000001, 0x2fa401ad, 0x00000b06, 0x00000000 },
    { 0x00000001, 0x2fa801ad, 0x00000b26, 0x00000000 },
    { 0x00000040, 0x2fe00c01, 0x00001400, 0x00000020 },
-   { 0x00000020, 0x34001c00, 0x00001400, 0x00000750 },
+   { 0x00000020, 0x34001c00, 0x00001400, 0x000007e0 },
    { 0x00000001, 0x2ac201ad, 0x00000fe4, 0x00000000 },
+   { 0x0020000c, 0x2a803dad, 0x00450ac0, 0x00020002 },
+   { 0x00200040, 0x2a883dad, 0x00450a80, 0x00030003 },
+   { 0x00200005, 0x2a902d29, 0x00450a88, 0xfffcfffc },
    { 0x00600001, 0x25600021, 0x008d0020, 0x00000000 },
    { 0x00600001, 0x28400021, 0x008d0560, 0x00000000 },
    { 0x00600001, 0x28600061, 0x00000000, 0x00000000 },
@@ -193,6 +197,12 @@
    { 0x00000001, 0x244001ed, 0x00000000, 0xfff0fff0 },
    { 0x00000001, 0x244201ed, 0x00000000, 0xfff4fff4 },
    { 0x00000001, 0x24440021, 0x00000440, 0x00000000 },
+   { 0x01000005, 0x20002e28, 0x000000a5, 0x00600060 },
+   { 0x00010040, 0x24403dad, 0x00000440, 0x000c000c },
+   { 0x01000005, 0x20002e28, 0x000000a5, 0x00100010 },
+   { 0x00010040, 0x24423dad, 0x00000442, 0x00080008 },
+   { 0x00200040, 0x244035ad, 0x00450440, 0x00450a90 },
+   { 0x00200040, 0x244435ad, 0x00450444, 0x00450a90 },
    { 0x00600001, 0x28000021, 0x008d0440, 0x00000000 },
    { 0x00000001, 0x24600061, 0x00000000, 0x00000002 },
    { 0x00000001, 0x24640231, 0x0000009c, 0x00000000 },
diff --git a/src/shaders/vme/vme75.inc b/src/shaders/vme/vme75.inc
index 5a121b9..1c286c0 100644
--- a/src/shaders/vme/vme75.inc
+++ b/src/shaders/vme/vme75.inc
@@ -309,6 +309,7 @@ define(`mb_mv1',		`r94')
 define(`mb_mv2',		`r95')
 define(`mb_mv3',		`r96')
 define(`mb_ref',		`r97')
+define(`mb_ref_win',		`r84')
 
 define(`DREF_REGION_SIZE',               `0x2020:UW')
 define(`PRED_L0',               `0x0':uw)