[965-xvmc] fix dual prime kernel, flush issue on G4x

3 files changed, 141 insertions, 178 deletions

diff --git a/src/xvmc/dual_prime.g4a b/src/xvmc/dual_prime.g4a
index f1a8c31b..7066a75b 100644
--- a/src/xvmc/dual_prime.g4a
+++ b/src/xvmc/dual_prime.g4a
@@ -13,11 +13,11 @@
  * Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDINg BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINgEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIgHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAgES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISINg FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINgS IN THE
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
  * Author:
@@ -26,23 +26,21 @@
  *    Xing Dong sheng <dongsheng.xing@intel.com>
  *
  */
-
-//and (1) g1.4<1>UD g1.4<1,1,1>UD 0xFFFFFFFEUD {align1};
 mov (8) g76.0<1>UD g1.0<8,8,1>UD {align1};
+//mov (8) g77.0<1>UD g2.0<8,8,1>UD {align1};
+
+include(`block_clear.g4i')
 mov (8) g115.0<1>UD g1.0<8,8,1>UD {align1};
 mov (8) g116.0<1>UD g1.0<8,8,1>UD {align1};
-mov(1) g115.8<1>UD 0x007001fUD  { align1 };
-//mov(1) g1.8<1>UD 0x007000fUD  { align1 };
 
-/*first vector---Y---top*/
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
+/* forward---Y---first vector*/
+mov(1) g115.8<1>UD 0x007001fUD  { align1 };
 asr (2) g115.14<1>W g1.14<2,2,1>W 1W {align1};
 add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-//add (1) g115.4<1>UD g115.4<1,1,1>UD 0UD {align1};
 mov(1) g115.8<1>UD 0x1fUD  { align1 }; //read 1 line, 32 columns.
+and.nz (1) null g1.31<1,1,1>UB 0x1UW {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
 define(`surface',`4')
 define(`mv1',`g1.14')
 define(`mv2',`g1.16')
@@ -56,17 +54,16 @@ mov (8) g62.0<1>UD g101.0<8,8,1>UD {align1};
 mov (8) g64.0<1>UD g102.0<8,8,1>UD {align1};
 mov (8) g66.0<1>UD g103.0<8,8,1>UD {align1};
 
-/*first vector---Y---bottom*/
-asr (2) g115.14<1>W g1.18<2,2,1>W 1W {align1};
-add (2) g115.0<1>UD g116.0<2,2,1>UD g115.14<2,2,1>W {align1};
+/*forward---Y---second vector*/
+asr (2) g115.14<1>W g1.22<2,2,1>W 1W {align1};
+add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 mov(1) g115.8<1>UD 0x1fUD { align1 }; //read 1 line, 32 columns.
+and.nz (1) null g1.31<1,1,1>UB 0x4UD {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
 define(`surface',`4')
-define(`mv1',`g1.18')
-define(`mv2',`g1.20')
+define(`mv1',`g1.22')
+define(`mv2',`g1.24')
 include(`motion_field_y.g4i')
 mov (8) g53.0<1>UD g96.0<8,8,1>UD {align1};
 mov (8) g55.0<1>UD g97.0<8,8,1>UD {align1};
@@ -77,14 +74,14 @@ mov (8) g63.0<1>UD g101.0<8,8,1>UD {align1};
 mov (8) g65.0<1>UD g102.0<8,8,1>UD {align1};
 mov (8) g67.0<1>UD g103.0<8,8,1>UD {align1};
 
-/*first vector---UV---top*/
+/*forward---UV---first vector*/
 mov(1) g115.8<1>UD 0x007000fUD  { align1 };
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.14<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
+asr (2) g115.14<1>W g1.14<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x1UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 define(`surface_u',`5')
 define(`surface_v',`6')
 define(`mv1',`g1.14')
@@ -99,19 +96,15 @@ mov (8) g73.0<1>UW g83.0<8,8,1>UW {align1};
 mov (8) g74.0<1>UW g84.0<8,8,1>UW {align1};
 mov (8) g75.0<1>UW g85.0<8,8,1>UW {align1};
 
-
-/*first vector---UV---bottom*/
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.18<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-define(`surface_u',`5')
-define(`surface_v',`6')
-define(`mv1',`g1.18')
-define(`mv2',`g1.20')
+/*forward---UV---second vector */
+asr (2) g115.14<1>W g1.22<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x4UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
+define(`mv1',`g1.24')
+define(`mv2',`g1.26')
 include(`motion_field_uv.g4i')
 mov (8) g68.16<1>UW g78.0<8,8,1>UW {align1};
 mov (8) g69.16<1>UW g79.0<8,8,1>UW {align1};
@@ -122,25 +115,19 @@ mov (8) g73.16<1>UW g83.0<8,8,1>UW {align1};
 mov (8) g74.16<1>UW g84.0<8,8,1>UW {align1};
 mov (8) g75.16<1>UW g85.0<8,8,1>UW {align1};
 
-
-
-/*second vector---Y---top*/
-//mov(1) g115.8<1>UD 0x007001fUD  { align1 };
-//mov(1) g1.8<1>UD 0x007000fUD  { align1 };
-//cmp.g (1) null g1.14<1,1,1>W 0W {align1};
-//(f0) add (1) g1.22<1>W g1.22<1,1,1>W 1W {align1};
-//cmp.g (1) null g1.16<1,1,1>W 0W {align1};
-//(f0) add (1) g1.24<1>W g1.24<1,1,1>W 10W {align1};
-asr (2) g115.14<1>W g1.22<2,2,1>W 1W {align1};
-add (2) g115.0<1>UD g116.0<2,2,1>UD g115.14<2,2,1>W {align1};
+/*backward---Y---first vector */
+mov(8) g1.0<1>UD g116.0<8,8,1>UD {align1};
+mov(1) g115.8<1>UD 0x007001fUD  { align1 };
+mov(1) g1.8<1>UD 0x007000fUD  { align1 };
+asr (2) g115.14<1>W g1.18<2,2,1>W 1W {align1};
+add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1) null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1D {align1};
 mov(1) g115.8<1>UD 0x1fUD  { align1 }; //read 1 line, 32 columns.
-define(`surface',`4')
-define(`mv1',`g1.22')
-define(`mv2',`g1.24')
+and.nz (1) null g1.31<1,1,1>UB 0x2UW {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
+define(`surface',`7')
+define(`mv1',`g1.18')
+define(`mv2',`g1.20')
 include(`motion_field_y.g4i')
 mov (8) g28.0<1>UD g96.0<8,8,1>UD {align1};
 mov (8) g30.0<1>UD g97.0<8,8,1>UD {align1};
@@ -151,18 +138,14 @@ mov (8) g38.0<1>UD g101.0<8,8,1>UD {align1};
 mov (8) g40.0<1>UD g102.0<8,8,1>UD {align1};
 mov (8) g42.0<1>UD g103.0<8,8,1>UD {align1};
 
-/*second vector---Y---bottom*/
-//cmp.g (1) null g1.18<1,1,1>W 0W {align1};
-//(f0) add (1) g1.26<1>W g1.26<1,1,1>W 1W {align1};
-cmp.g (1) null g1.20<1,1,1>W 0W {align1};
-(f0) add (1) g1.28<1>W g1.28<1,1,1>W 5W {align1};
+/*backward---Y---second vector */
 asr (2) g115.14<1>W g1.26<2,2,1>W 1W {align1};
-add (2) g115.0<1>UD g116.0<2,2,1>UD g115.14<2,2,1>W {align1};
+add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1)null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 mov(1) g115.8<1>UD 0x1fUD  { align1 }; //read 1 line, 32 columns.
-define(`surface',`4')
+and.nz (1) null g2.20<1,1,1>UD 0x8UD {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
+define(`surface',`7')
 define(`mv1',`g1.26')
 define(`mv2',`g1.28')
 include(`motion_field_y.g4i')
@@ -184,19 +167,18 @@ avg.sat (16) g41.0<1>UW g65.0<16,16,1>UW g102.0<16,16,1>UW {align1};
 avg.sat (16) g42.0<1>UW g66.0<16,16,1>UW g42.0<16,16,1>UW {align1};
 avg.sat (16) g43.0<1>UW g67.0<16,16,1>UW g103.0<16,16,1>UW {align1};
 
-/*second vector---UV---top*/
+/*backward---UV---first vector */
 mov(1) g115.8<1>UD 0x007000fUD  { align1 };
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.22<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-define(`surface_u',`5')
-define(`surface_v',`6')
-define(`mv1',`g1.22')
-define(`mv2',`g1.24')
+asr (2) g115.14<1>W g1.18<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x2UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
+define(`surface_u',`8')
+define(`surface_v',`9')
+define(`mv1',`g1.18')
+define(`mv2',`g1.20')
 include(`motion_field_uv.g4i')
 mov (8) g44.0<1>UW g78.0<8,8,1>UW {align1};
 mov (8) g45.0<1>UW g79.0<8,8,1>UW {align1};
@@ -207,15 +189,13 @@ mov (8) g49.0<1>UW g83.0<8,8,1>UW {align1};
 mov (8) g50.0<1>UW g84.0<8,8,1>UW {align1};
 mov (8) g51.0<1>UW g85.0<8,8,1>UW {align1};
 
-/*second vector---UV---bottom*/
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.26<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-define(`surface_u',`5')
-define(`surface_v',`6')
+/*backward---UV---second vector */
+asr (2) g115.14<1>W g1.26<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x8UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 define(`mv1',`g1.26')
 define(`mv2',`g1.28')
 include(`motion_field_uv.g4i')
@@ -236,6 +216,7 @@ avg.sat (16) g48.0<1>UW g72.0<16,16,1>UW g48.0<16,16,1>UW {align1};
 avg.sat (16) g49.0<1>UW g73.0<16,16,1>UW g49.0<16,16,1>UW {align1};
 avg.sat (16) g50.0<1>UW g74.0<16,16,1>UW g50.0<16,16,1>UW {align1};
 avg.sat (16) g51.0<1>UW g75.0<16,16,1>UW g51.0<16,16,1>UW {align1};
+
 include(`addidct.g4i')
 
 //send (16) 0 acc0<1>UW g0<8,8,1>UW 
diff --git a/src/xvmc/dual_prime_igd.g4a b/src/xvmc/dual_prime_igd.g4a
index 4a59cadd..e741244b 100644
--- a/src/xvmc/dual_prime_igd.g4a
+++ b/src/xvmc/dual_prime_igd.g4a
@@ -13,11 +13,11 @@
  * Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDINg BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINgEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIgHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAgES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISINg FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINgS IN THE
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  *
  * Author:
@@ -26,23 +26,21 @@
  *    Xing Dong sheng <dongsheng.xing@intel.com>
  *
  */
-
-//and (1) g1.4<1>UD g1.4<1,1,1>UD 0xFFFFFFFEUD {align1};
 mov (8) g76.0<1>UD g1.0<8,8,1>UD {align1};
+//mov (8) g77.0<1>UD g2.0<8,8,1>UD {align1};
+
+include(`block_clear.g4i')
 mov (8) g115.0<1>UD g1.0<8,8,1>UD {align1};
 mov (8) g116.0<1>UD g1.0<8,8,1>UD {align1};
-mov(1) g115.8<1>UD 0x007001fUD  { align1 };
-//mov(1) g1.8<1>UD 0x007000fUD  { align1 };
 
-/*first vector---Y---top*/
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
+/* forward---Y---first vector*/
+mov(1) g115.8<1>UD 0x007001fUD  { align1 };
 asr (2) g115.14<1>W g1.14<2,2,1>W 1W {align1};
 add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-//add (1) g115.4<1>UD g115.4<1,1,1>UD 0UD {align1};
 mov(1) g115.8<1>UD 0x1fUD  { align1 }; //read 1 line, 32 columns.
+and.nz (1) null g1.31<1,1,1>UB 0x1UW {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
 define(`surface',`4')
 define(`mv1',`g1.14')
 define(`mv2',`g1.16')
@@ -56,17 +54,16 @@ mov (8) g62.0<1>UD g101.0<8,8,1>UD {align1};
 mov (8) g64.0<1>UD g102.0<8,8,1>UD {align1};
 mov (8) g66.0<1>UD g103.0<8,8,1>UD {align1};
 
-/*first vector---Y---bottom*/
-asr (2) g115.14<1>W g1.18<2,2,1>W 1W {align1};
-add (2) g115.0<1>UD g116.0<2,2,1>UD g115.14<2,2,1>W {align1};
+/*forward---Y---second vector*/
+asr (2) g115.14<1>W g1.22<2,2,1>W 1W {align1};
+add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 mov(1) g115.8<1>UD 0x1fUD { align1 }; //read 1 line, 32 columns.
+and.nz (1) null g1.31<1,1,1>UB 0x4UD {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
 define(`surface',`4')
-define(`mv1',`g1.18')
-define(`mv2',`g1.20')
+define(`mv1',`g1.22')
+define(`mv2',`g1.24')
 include(`motion_field_y_igd.g4i')
 mov (8) g53.0<1>UD g96.0<8,8,1>UD {align1};
 mov (8) g55.0<1>UD g97.0<8,8,1>UD {align1};
@@ -77,14 +74,14 @@ mov (8) g63.0<1>UD g101.0<8,8,1>UD {align1};
 mov (8) g65.0<1>UD g102.0<8,8,1>UD {align1};
 mov (8) g67.0<1>UD g103.0<8,8,1>UD {align1};
 
-/*first vector---UV---top*/
+/*forward---UV---first vector*/
 mov(1) g115.8<1>UD 0x007000fUD  { align1 };
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.14<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
+asr (2) g115.14<1>W g1.14<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x1UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 define(`surface_u',`5')
 define(`surface_v',`6')
 define(`mv1',`g1.14')
@@ -99,19 +96,15 @@ mov (8) g73.0<1>UW g83.0<8,8,1>UW {align1};
 mov (8) g74.0<1>UW g84.0<8,8,1>UW {align1};
 mov (8) g75.0<1>UW g85.0<8,8,1>UW {align1};
 
-
-/*first vector---UV---bottom*/
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.18<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 1UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-define(`surface_u',`5')
-define(`surface_v',`6')
-define(`mv1',`g1.18')
-define(`mv2',`g1.20')
+/*forward---UV---second vector */
+asr (2) g115.14<1>W g1.22<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x4UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
+define(`mv1',`g1.24')
+define(`mv2',`g1.26')
 include(`motion_field_uv_igd.g4i')
 mov (8) g68.16<1>UW g78.0<8,8,1>UW {align1};
 mov (8) g69.16<1>UW g79.0<8,8,1>UW {align1};
@@ -122,25 +115,19 @@ mov (8) g73.16<1>UW g83.0<8,8,1>UW {align1};
 mov (8) g74.16<1>UW g84.0<8,8,1>UW {align1};
 mov (8) g75.16<1>UW g85.0<8,8,1>UW {align1};
 
-
-
-/*second vector---Y---top*/
-//mov(1) g115.8<1>UD 0x007001fUD  { align1 };
-//mov(1) g1.8<1>UD 0x007000fUD  { align1 };
-//cmp.g (1) null g1.14<1,1,1>W 0W {align1};
-//(f0) add (1) g1.22<1>W g1.22<1,1,1>W 1W {align1};
-//cmp.g (1) null g1.16<1,1,1>W 0W {align1};
-//(f0) add (1) g1.24<1>W g1.24<1,1,1>W 10W {align1};
-asr (2) g115.14<1>W g1.22<2,2,1>W 1W {align1};
-add (2) g115.0<1>UD g116.0<2,2,1>UD g115.14<2,2,1>W {align1};
+/*backward---Y---first vector */
+mov(8) g1.0<1>UD g116.0<8,8,1>UD {align1};
+mov(1) g115.8<1>UD 0x007001fUD  { align1 };
+mov(1) g1.8<1>UD 0x007000fUD  { align1 };
+asr (2) g115.14<1>W g1.18<2,2,1>W 1W {align1};
+add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1) null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1D {align1};
 mov(1) g115.8<1>UD 0x1fUD  { align1 }; //read 1 line, 32 columns.
-define(`surface',`4')
-define(`mv1',`g1.22')
-define(`mv2',`g1.24')
+and.nz (1) null g1.31<1,1,1>UB 0x2UW {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
+define(`surface',`7')
+define(`mv1',`g1.18')
+define(`mv2',`g1.20')
 include(`motion_field_y_igd.g4i')
 mov (8) g28.0<1>UD g96.0<8,8,1>UD {align1};
 mov (8) g30.0<1>UD g97.0<8,8,1>UD {align1};
@@ -151,18 +138,14 @@ mov (8) g38.0<1>UD g101.0<8,8,1>UD {align1};
 mov (8) g40.0<1>UD g102.0<8,8,1>UD {align1};
 mov (8) g42.0<1>UD g103.0<8,8,1>UD {align1};
 
-/*second vector---Y---bottom*/
-//cmp.g (1) null g1.18<1,1,1>W 0W {align1};
-//(f0) add (1) g1.26<1>W g1.26<1,1,1>W 1W {align1};
-cmp.g (1) null g1.20<1,1,1>W 0W {align1};
-(f0) add (1) g1.28<1>W g1.28<1,1,1>W 5W {align1};
+/*backward---Y---second vector */
 asr (2) g115.14<1>W g1.26<2,2,1>W 1W {align1};
-add (2) g115.0<1>UD g116.0<2,2,1>UD g115.14<2,2,1>W {align1};
+add (2) g115.0<1>D g116.0<2,2,1>D g115.14<2,2,1>W {align1};
 and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
-//and.nz (1)null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 mov(1) g115.8<1>UD 0x1fUD  { align1 }; //read 1 line, 32 columns.
-define(`surface',`4')
+and.nz (1) null g2.20<1,1,1>UD 0x8UD {align1};
+(f0) add (1) g115.4<1>D g115.4<1,1,1>D 1D {align1};
+define(`surface',`7')
 define(`mv1',`g1.26')
 define(`mv2',`g1.28')
 include(`motion_field_y_igd.g4i')
@@ -184,19 +167,18 @@ avg.sat (16) g41.0<1>UW g65.0<16,16,1>UW g102.0<16,16,1>UW {align1};
 avg.sat (16) g42.0<1>UW g66.0<16,16,1>UW g42.0<16,16,1>UW {align1};
 avg.sat (16) g43.0<1>UW g67.0<16,16,1>UW g103.0<16,16,1>UW {align1};
 
-/*second vector---UV---top*/
+/*backward---UV---first vector */
 mov(1) g115.8<1>UD 0x007000fUD  { align1 };
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.22<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-define(`surface_u',`5')
-define(`surface_v',`6')
-define(`mv1',`g1.22')
-define(`mv2',`g1.24')
+asr (2) g115.14<1>W g1.18<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x2UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
+define(`surface_u',`8')
+define(`surface_v',`9')
+define(`mv1',`g1.18')
+define(`mv2',`g1.20')
 include(`motion_field_uv_igd.g4i')
 mov (8) g44.0<1>UW g78.0<8,8,1>UW {align1};
 mov (8) g45.0<1>UW g79.0<8,8,1>UW {align1};
@@ -207,15 +189,13 @@ mov (8) g49.0<1>UW g83.0<8,8,1>UW {align1};
 mov (8) g50.0<1>UW g84.0<8,8,1>UW {align1};
 mov (8) g51.0<1>UW g85.0<8,8,1>UW {align1};
 
-/*second vector---UV---bottom*/
-asr (2) g115.0<1>UD g116.0<1,1,1>UD 1UD {align1}; // x/=2 y/=2 
-asr (2) g115.14<1>W g1.26<2,2,1>W 2W {align1};   // vector/=4
-add (2) g115.0<1>UD g115.0<2,2,1>UD g115.14<2,2,1>W {align1}; // (x,y)+=vector
-and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};  // cut y
-//and.nz (1) null g1.31<1,1,1>UB 4UW {align1};
-//(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
-define(`surface_u',`5')
-define(`surface_v',`6')
+/*backward---UV---second vector */
+asr (2) g115.14<1>W g1.26<2,2,1>W 2W {align1};
+asr (2) g115.0<1>D g116.0<2,2,1>D 1D {align1};
+add (2) g115.0<1>D g115.0<2,2,1>D g115.14<2,2,1>W {align1};
+and (1) g115.4<1>UD g115.4<1,1,1>UD 0xFFFFFFFEUD {align1};
+and.nz (1) null g1.31<1,1,1>UB 0x8UW {align1};
+(f0) add (1) g115.4<1>UD g115.4<1,1,1>UD 1UD {align1};
 define(`mv1',`g1.26')
 define(`mv2',`g1.28')
 include(`motion_field_uv_igd.g4i')
@@ -236,6 +216,7 @@ avg.sat (16) g48.0<1>UW g72.0<16,16,1>UW g48.0<16,16,1>UW {align1};
 avg.sat (16) g49.0<1>UW g73.0<16,16,1>UW g49.0<16,16,1>UW {align1};
 avg.sat (16) g50.0<1>UW g74.0<16,16,1>UW g50.0<16,16,1>UW {align1};
 avg.sat (16) g51.0<1>UW g75.0<16,16,1>UW g51.0<16,16,1>UW {align1};
+
 include(`addidct.g4i')
 
 //send (16) 0 acc0<1>UW g0<8,8,1>UW 
diff --git a/src/xvmc/i965_xvmc.c b/src/xvmc/i965_xvmc.c
index 8ccdd775..31d68969 100644
--- a/src/xvmc/i965_xvmc.c
+++ b/src/xvmc/i965_xvmc.c
@@ -176,6 +176,7 @@ static void flush()
     struct brw_mi_flush flush;
     memset(&flush, 0, sizeof(flush));
     flush.opcode = CMD_MI_FLUSH;
+    flush.flags = (1<<1);
     BATCH_STRUCT(flush);
 }