VPP: MADI on SNB

Set the right surface states for reference, STMM and output surface,
fix the shader as well

Signed-off-by: Xiang, Haihao <haihao.xiang@intel.com>
Tested-By: Simon Farnsworth <simon.farnsworth@onelan.co.uk>
(cherry picked from commit 1d1b8da1284f7f918733db79428f09af38d7e14a)

Conflicts:

	src/i965_post_processing.c

5 files changed, 203 insertions, 86 deletions

diff --git a/src/i965_device_info.c b/src/i965_device_info.c
index 4fad7a4..6fad106 100644
--- a/src/i965_device_info.c
+++ b/src/i965_device_info.c
@@ -104,6 +104,7 @@ static const struct hw_codec_info snb_hw_codec_info = {
     .has_accelerated_getimage = 1,
     .has_accelerated_putimage = 1,
     .has_tiled_surface = 1,
+    .has_di_motion_adptive = 1,
 
     .num_filters = 2,
     .filters = {
diff --git a/src/i965_post_processing.c b/src/i965_post_processing.c
index c2db480..acc9732 100755
--- a/src/i965_post_processing.c
+++ b/src/i965_post_processing.c
@@ -40,6 +40,11 @@
 #include "i965_render.h"
 #include "intel_media.h"
 
+extern VAStatus
+vpp_surface_convert(VADriverContextP ctx,
+                    struct object_surface *src_obj_surf,
+                    struct object_surface *dst_obj_surf);
+
 #define HAS_VPP(ctx) ((ctx)->codec_info->has_vpp)
 
 #define SURFACE_STATE_PADDED_SIZE               MAX(SURFACE_STATE_PADDED_SIZE_GEN8,\
@@ -3002,13 +3007,14 @@ pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_contex
     struct pp_dndi_context *pp_dndi_context = (struct pp_dndi_context *)&pp_context->pp_dndi_context;
     struct pp_inline_parameter *pp_inline_parameter = pp_context->pp_inline_parameter;
     struct pp_static_parameter *pp_static_parameter = pp_context->pp_static_parameter;
-    struct object_surface *obj_surface;
+    struct object_surface *previous_in_obj_surface, *current_in_obj_surface, *previous_out_obj_surface, *current_out_obj_surface;
     struct i965_sampler_dndi *sampler_dndi;
     int index;
     int w, h;
     int orig_w, orig_h;
     int dndi_top_first = 1;
     VAProcFilterParameterBufferDeinterlacing *di_filter_param = (VAProcFilterParameterBufferDeinterlacing *)filter_param;
+    int is_first_frame = (pp_dndi_context->frame_order == -1);
 
     if (di_filter_param->flags & VA_DEINTERLACING_BOTTOM_FIELD)
         dndi_top_first = 0;
@@ -3016,58 +3022,159 @@ pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_contex
         dndi_top_first = 1;
 
     /* surface */
-    obj_surface = (struct object_surface *)src_surface->base;
-    orig_w = obj_surface->orig_width;
-    orig_h = obj_surface->orig_height;
-    w = obj_surface->width;
-    h = obj_surface->height;
+    current_in_obj_surface = (struct object_surface *)src_surface->base;
 
-    if (pp_dndi_context->stmm_bo == NULL) {
-        pp_dndi_context->stmm_bo = dri_bo_alloc(i965->intel.bufmgr,
-                                                "STMM surface",
-                                                w * h,
-                                                4096);
-        assert(pp_dndi_context->stmm_bo);
+    if (di_filter_param->algorithm == VAProcDeinterlacingBob) {
+        previous_in_obj_surface = current_in_obj_surface;
+        is_first_frame = 1;
+    } else if (di_filter_param->algorithm == VAProcDeinterlacingMotionAdaptive) {
+        if (pp_dndi_context->frame_order == 0) {
+            VAProcPipelineParameterBuffer *pipeline_param = pp_context->pipeline_param;
+            if (!pipeline_param ||
+                !pipeline_param->num_forward_references ||
+                pipeline_param->forward_references[0] == VA_INVALID_ID) {
+                WARN_ONCE("A forward temporal reference is needed for Motion adaptive deinterlacing !!!\n");
+
+                return VA_STATUS_ERROR_INVALID_PARAMETER;
+            } else {
+                previous_in_obj_surface = SURFACE(pipeline_param->forward_references[0]);
+                assert(previous_in_obj_surface && previous_in_obj_surface->bo);
+
+                is_first_frame = 0;
+            }
+        } else if (pp_dndi_context->frame_order == 1) {
+            vpp_surface_convert(ctx,
+                                pp_dndi_context->current_out_obj_surface,
+                                (struct object_surface *)dst_surface->base);
+            pp_dndi_context->frame_order = (pp_dndi_context->frame_order + 1) % 2;
+            is_first_frame = 0;
+
+            return VA_STATUS_SUCCESS_1;
+        } else {
+            previous_in_obj_surface = current_in_obj_surface;
+            is_first_frame = 1;
+        }
+    } else {
+        return VA_STATUS_ERROR_UNIMPLEMENTED;
     }
 
+    /* source (temporal reference) YUV surface index 5 */
+    orig_w = previous_in_obj_surface->orig_width;
+    orig_h = previous_in_obj_surface->orig_height;
+    w = previous_in_obj_surface->width;
+    h = previous_in_obj_surface->height;
+    i965_pp_set_surface2_state(ctx, pp_context,
+                               previous_in_obj_surface->bo, 0,
+                               orig_w, orig_h, w,
+                               0, h,
+                               SURFACE_FORMAT_PLANAR_420_8, 1,
+                               5);
+
+    /* source surface */
+    orig_w = current_in_obj_surface->orig_width;
+    orig_h = current_in_obj_surface->orig_height;
+    w = current_in_obj_surface->width;
+    h = current_in_obj_surface->height;
+
     /* source UV surface index 2 */
     i965_pp_set_surface_state(ctx, pp_context,
-                              obj_surface->bo, w * h,
+                              current_in_obj_surface->bo, w * h,
                               orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM,
                               2, 0);
 
     /* source YUV surface index 4 */
     i965_pp_set_surface2_state(ctx, pp_context,
-                               obj_surface->bo, 0,
+                               current_in_obj_surface->bo, 0,
                                orig_w, orig_h, w,
                                0, h,
                                SURFACE_FORMAT_PLANAR_420_8, 1,
                                4);
 
-    /* source STMM surface index 20 */
+    /* source STMM surface index 6 */
+    if (pp_dndi_context->stmm_bo == NULL) {
+        pp_dndi_context->stmm_bo = dri_bo_alloc(i965->intel.bufmgr,
+                                                "STMM surface",
+                                                w * h,
+                                                4096);
+        assert(pp_dndi_context->stmm_bo);
+    }
+
     i965_pp_set_surface_state(ctx, pp_context,
                               pp_dndi_context->stmm_bo, 0,
                               orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM,
-                              20, 1);
+                              6, 0);
 
-    /* destination surface */
-    obj_surface = (struct object_surface *)dst_surface->base;
-    orig_w = obj_surface->orig_width;
-    orig_h = obj_surface->orig_height;
-    w = obj_surface->width;
-    h = obj_surface->height;
+    /* destination (Previous frame) */
+    previous_out_obj_surface = (struct object_surface *)dst_surface->base;
+    orig_w = previous_out_obj_surface->orig_width;
+    orig_h = previous_out_obj_surface->orig_height;
+    w = previous_out_obj_surface->width;
+    h = previous_out_obj_surface->height;
 
-    /* destination Y surface index 7 */
+    if (is_first_frame) {
+        current_out_obj_surface = previous_out_obj_surface;
+    } else {
+        VAStatus va_status;
+
+        if (pp_dndi_context->current_out_surface == VA_INVALID_SURFACE) {
+            unsigned int tiling = 0, swizzle = 0;
+            dri_bo_get_tiling(previous_out_obj_surface->bo, &tiling, &swizzle);
+
+            va_status = i965_CreateSurfaces(ctx,
+                                            orig_w,
+                                            orig_h,
+                                            VA_RT_FORMAT_YUV420,
+                                            1,
+                                            &pp_dndi_context->current_out_surface);
+            assert(va_status == VA_STATUS_SUCCESS);
+            pp_dndi_context->current_out_obj_surface = SURFACE(pp_dndi_context->current_out_surface);
+            assert(pp_dndi_context->current_out_obj_surface);
+            i965_check_alloc_surface_bo(ctx,
+                                        pp_dndi_context->current_out_obj_surface,
+                                        tiling != I915_TILING_NONE,
+                                        VA_FOURCC_NV12,
+                                        SUBSAMPLE_YUV420);
+        }
+
+        current_out_obj_surface = pp_dndi_context->current_out_obj_surface;
+    }
+
+    /* destination (Previous frame) Y surface index 7 */
     i965_pp_set_surface_state(ctx, pp_context,
-                              obj_surface->bo, 0,
+                              previous_out_obj_surface->bo, 0,
                               orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM,
                               7, 1);
 
-    /* destination UV surface index 8 */
+    /* destination (Previous frame) UV surface index 8 */
     i965_pp_set_surface_state(ctx, pp_context,
-                              obj_surface->bo, w * h,
+                              previous_out_obj_surface->bo, w * h,
                               orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM,
                               8, 1);
+
+    /* destination(Current frame) */
+    orig_w = current_out_obj_surface->orig_width;
+    orig_h = current_out_obj_surface->orig_height;
+    w = current_out_obj_surface->width;
+    h = current_out_obj_surface->height;
+
+    /* destination (Current frame) Y surface index xxx */
+    i965_pp_set_surface_state(ctx, pp_context,
+                              current_out_obj_surface->bo, 0,
+                              orig_w / 4, orig_h, w, I965_SURFACEFORMAT_R8_UNORM,
+                              10, 1);
+
+    /* destination (Current frame) UV surface index xxx */
+    i965_pp_set_surface_state(ctx, pp_context,
+                              current_out_obj_surface->bo, w * h,
+                              orig_w / 4, orig_h / 2, w, I965_SURFACEFORMAT_R8G8_UNORM,
+                              11, 1);
+
+    /* STMM output surface, index 20 */
+    i965_pp_set_surface_state(ctx, pp_context,
+                              pp_dndi_context->stmm_bo, 0,
+                              orig_w, orig_h, w, I965_SURFACEFORMAT_R8_UNORM,
+                              20, 1);
+
     /* sampler dndi */
     dri_bo_map(pp_context->sampler_state_table.bo, True);
     assert(pp_context->sampler_state_table.bo->virtual);
@@ -3115,7 +3222,7 @@ pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_contex
     sampler_dndi[index].dw6.di_partial = 0;
     sampler_dndi[index].dw6.dndi_top_first = dndi_top_first;
     sampler_dndi[index].dw6.dndi_stream_id = 0;
-    sampler_dndi[index].dw6.dndi_first_frame = 1;
+    sampler_dndi[index].dw6.dndi_first_frame = is_first_frame;
     sampler_dndi[index].dw6.progressive_dn = 0;
     sampler_dndi[index].dw6.fmd_tear_threshold = 2;
     sampler_dndi[index].dw6.fmd2_vertical_difference_threshold = 100;
@@ -3149,6 +3256,8 @@ pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_contex
 
     dst_surface->flags = I965_SURFACE_FLAG_FRAME;
 
+    pp_dndi_context->frame_order = (pp_dndi_context->frame_order + 1) % 2;
+
     return VA_STATUS_SUCCESS;
 }
 
@@ -3386,12 +3495,6 @@ gen7_pp_dndi_set_block_parameter(struct i965_post_processing_context *pp_context
     return 0;
 }
 
-
-extern VAStatus
-vpp_surface_convert(VADriverContextP ctx,
-                    struct object_surface *src_obj_surf,
-                    struct object_surface *dst_obj_surf);
-
 static VAStatus
 gen7_pp_nv12_dndi_initialize(VADriverContextP ctx, struct i965_post_processing_context *pp_context,
                              const struct i965_surface *src_surface,
@@ -5283,8 +5386,6 @@ static const int proc_frame_to_pp_frame[3] = {
     I965_SURFACE_FLAG_BOTTOME_FIELD_FIRST
 };
 
-#define VA_STATUS_SUCCESS_1                     0xFFFFFFFE
-
 VAStatus 
 i965_proc_picture(VADriverContextP ctx, 
                   VAProfile profile, 
diff --git a/src/shaders/post_processing/gen5_6/Core_Kernels/PL_DNDI_ALG_UVCopy_NV12.asm b/src/shaders/post_processing/gen5_6/Core_Kernels/PL_DNDI_ALG_UVCopy_NV12.asm
index 280d37a..23bd306 100644
--- a/src/shaders/post_processing/gen5_6/Core_Kernels/PL_DNDI_ALG_UVCopy_NV12.asm
+++ b/src/shaders/post_processing/gen5_6/Core_Kernels/PL_DNDI_ALG_UVCopy_NV12.asm
@@ -62,24 +62,29 @@
     #include "DI_Hist_Save.asm"
 
 ////////////////////////////////////// Save the DN Curr Frame for Next Run ////////////////////////
-    add (4)     pCF_Y_OFFSET<1>:uw          ubSRC_CF_OFFSET<4;4,1>:ub  npDN_YUV:w
-    // check top/bottom field first
-    cmp.e.f0.0 (1)  null<1>:w               ubTFLD_FIRST<0;1,0>:ub     1:w
-    (f0.0) jmpi (1) TOP_FIELD_FIRST
-
-BOTTOM_FIELD_FIRST:
-    $for (0,0; <nY_NUM_OF_ROWS/2; 2,1) {
-        mov (4)     mudMSGHDR_DN(1,%1*4)<1>     udRESP(nDI_CURR_2ND_FIELD_LUMA_OFFSET,%2*4)<4;4,1> // 2nd field luma from current frame (line 0,2)
-        mov (4)     mudMSGHDR_DN(1,%1*4+4)<1>   udRESP(nDI_CURR_FRAME_LUMA_OFFSET+%2,4)<4;4,1> // 1st field luma from current frame (line 1,3)
-    }
-    jmpi (1) SAVE_DN_CURR
-    
-TOP_FIELD_FIRST:
-    $for (0,0; <nY_NUM_OF_ROWS/2; 2,1) {
-        mov (4)     mudMSGHDR_DN(1,%1*4)<1>     udRESP(nDI_CURR_FRAME_LUMA_OFFSET+%2,0)<4;4,1> // 2nd field luma from current frame (line 0,2)
-        mov (4)     mudMSGHDR_DN(1,%1*4+4)<1>   udRESP(nDI_CURR_2ND_FIELD_LUMA_OFFSET,%2*4)<4;4,1> // 1st field luma from current frame (line 1,3)
+    // previous frame
+    $for (0; <nY_NUM_OF_ROWS/2; 1) {
+        mov (16) mubMSGHDR_DN(1, %1*16)<1>       ubRESP(nDI_PREV_FRAME_LUMA_OFFSET,%1*16)
     }
-SAVE_DN_CURR:
+
+    mov (2)     rMSGSRC.0<1>:ud        wORIX<2;2,1>:w               // X origin and Y origin
+    mov (1)     rMSGSRC.2<1>:ud        nDPW_BLOCK_SIZE_DN:ud        // block width and height (16x4)
+    mov (8)     mudMSGHDR_DN(0)<1>     rMSGSRC.0<8;8,1>:ud
+    send (8)    dNULLREG    mMSGHDR_DN   udDUMMY_NULL    nDATAPORT_WRITE    nDPMW_MSGDSC+nDPMW_MSG_LEN_PL_DN_DI+nBI_DESTINATION_Y:ud
+
+    //Write UV through DATAPORT
+    mov (2)     rMSGSRC.0<1>:ud        wORIX<2;2,1>:w               // X origin and Y origin
+    asr (1)     rMSGSRC.1<1>:d         rMSGSRC.1<0;1,0>:d    1:w  // U/V block origin should be half of Y's
+    mov (1)     rMSGSRC.2<1>:ud        nDPR_BLOCK_SIZE_UV:ud        // block width and height (16x2)
+    mov (8)     mudMSGHDR_DN(0)<1>     rMSGSRC.0<8;8,1>:ud
+
+    mov (8) mubMSGHDR_DN(1, 0)<2>      ubRESP(nDI_PREV_FRAME_CHROMA_OFFSET, 1)<16 ;8,2>
+    mov (8) mubMSGHDR_DN(1, 1)<2>      ubRESP(nDI_PREV_FRAME_CHROMA_OFFSET, 16)<16 ;8,2>
+    mov (8) mubMSGHDR_DN(1, 16)<2>      ubRESP(nDI_PREV_FRAME_CHROMA_OFFSET+1, 1)<16 ;8,2>
+    mov (8) mubMSGHDR_DN(1, 17)<2>      ubRESP(nDI_PREV_FRAME_CHROMA_OFFSET+1, 16)<16 ;8,2>
+    send (8)    dNULLREG    mMSGHDR_DN   udDUMMY_NULL    nDATAPORT_WRITE    nDPMW_MSGDSC+nMSGLEN_1+nBI_DESTINATION_UV:ud
+
+    // current frame
     $for (0; <nY_NUM_OF_ROWS/2; 1) {
         mov (16) mubMSGHDR_DN(1, %1*16)<1>       ubRESP(nDI_CURR_FRAME_LUMA_OFFSET,%1*16)
     }
@@ -87,7 +92,7 @@ SAVE_DN_CURR:
     mov (2)     rMSGSRC.0<1>:ud        wORIX<2;2,1>:w               // X origin and Y origin
     mov (1)     rMSGSRC.2<1>:ud        nDPW_BLOCK_SIZE_DN:ud        // block width and height (16x4)
     mov (8)     mudMSGHDR_DN(0)<1>     rMSGSRC.0<8;8,1>:ud
-    send (8)    dNULLREG    mMSGHDR_DN   udDUMMY_NULL    nDATAPORT_WRITE    nDPMW_MSGDSC+nDPMW_MSG_LEN_PL_DN_DI+nBI_DESTINATION_Y:ud
+    send (8)    dNULLREG    mMSGHDR_DN   udDUMMY_NULL    nDATAPORT_WRITE    nDPMW_MSGDSC+nDPMW_MSG_LEN_PL_DN_DI+nBI_DESTINATION_1_Y:ud
 
  		//Write UV through DATAPORT
 		mov (2)     rMSGSRC.0<1>:ud        wORIX<2;2,1>:w               // X origin and Y origin
@@ -99,4 +104,4 @@ SAVE_DN_CURR:
     mov (8) mubMSGHDR_DN(1, 1)<2>      ubRESP(nDI_CURR_FRAME_CHROMA_OFFSET, 16)<16 ;8,2>        
     mov (8) mubMSGHDR_DN(1, 16)<2>      ubRESP(nDI_CURR_FRAME_CHROMA_OFFSET+1, 1)<16 ;8,2>
     mov (8) mubMSGHDR_DN(1, 17)<2>      ubRESP(nDI_CURR_FRAME_CHROMA_OFFSET+1, 16)<16 ;8,2>        
-    send (8)    dNULLREG    mMSGHDR_DN   udDUMMY_NULL    nDATAPORT_WRITE    nDPMW_MSGDSC+nMSGLEN_1+nBI_DESTINATION_UV:ud 
\ No newline at end of file
+    send (8)    dNULLREG    mMSGHDR_DN   udDUMMY_NULL    nDATAPORT_WRITE    nDPMW_MSGDSC+nMSGLEN_1+nBI_DESTINATION_1_UV:ud
diff --git a/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g4b.gen5 b/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g4b.gen5
index 446fb4b..4563d20 100644
--- a/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g4b.gen5
+++ b/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g4b.gen5
@@ -44,18 +44,23 @@
    { 0x00600001, 0x21a00022, 0x008d0100, 0x00000000 },
    { 0x00000001, 0x21c00022, 0x00000560, 0x00000000 },
    { 0x0d600031, 0x20000c04, 0x508d0000, 0x04082014 },
-   { 0x00400040, 0x22083e28, 0x00690024, 0x07000700 },
-   { 0x01000010, 0x20003e2c, 0x0000003b, 0x00010001 },
-   { 0x00010220, 0x34001c00, 0x00001400, 0x0000000a },
-   { 0x00400001, 0x20400022, 0x00690580, 0x00000000 },
-   { 0x00400001, 0x20500022, 0x006904d0, 0x00000000 },
-   { 0x00400001, 0x20600022, 0x00690590, 0x00000000 },
-   { 0x00400001, 0x20700022, 0x006904f0, 0x00000000 },
-   { 0x00000220, 0x34001c00, 0x00001400, 0x00000008 },
-   { 0x00400001, 0x20400022, 0x006904c0, 0x00000000 },
-   { 0x00400001, 0x20500022, 0x00690580, 0x00000000 },
-   { 0x00400001, 0x20600022, 0x006904e0, 0x00000000 },
-   { 0x00400001, 0x20700022, 0x00690590, 0x00000000 },
+   { 0x00800001, 0x20400232, 0x00b10440, 0x00000000 },
+   { 0x00800001, 0x20500232, 0x00b10450, 0x00000000 },
+   { 0x00800001, 0x20600232, 0x00b10460, 0x00000000 },
+   { 0x00800001, 0x20700232, 0x00b10470, 0x00000000 },
+   { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
+   { 0x00000001, 0x21080061, 0x00000000, 0x0003000f },
+   { 0x00600001, 0x20200022, 0x008d0100, 0x00000000 },
+   { 0x01600031, 0x20000c04, 0x508d0000, 0x06082007 },
+   { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
+   { 0x0000000c, 0x21043ca5, 0x00000104, 0x00010001 },
+   { 0x00000001, 0x21080061, 0x00000000, 0x0001000f },
+   { 0x00600001, 0x20200022, 0x008d0100, 0x00000000 },
+   { 0x00600001, 0x40400232, 0x00ae0481, 0x00000000 },
+   { 0x00600001, 0x40410232, 0x00ae0490, 0x00000000 },
+   { 0x00600001, 0x40500232, 0x00ae04a1, 0x00000000 },
+   { 0x00600001, 0x40510232, 0x00ae04b0, 0x00000000 },
+   { 0x01600031, 0x20000c04, 0x508d0000, 0x04082008 },
    { 0x00800001, 0x20400232, 0x00b104c0, 0x00000000 },
    { 0x00800001, 0x20500232, 0x00b104d0, 0x00000000 },
    { 0x00800001, 0x20600232, 0x00b104e0, 0x00000000 },
@@ -63,7 +68,7 @@
    { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
    { 0x00000001, 0x21080061, 0x00000000, 0x0003000f },
    { 0x00600001, 0x20200022, 0x008d0100, 0x00000000 },
-   { 0x01600031, 0x20000c04, 0x508d0000, 0x06082007 },
+   { 0x01600031, 0x20000c04, 0x508d0000, 0x0608200a },
    { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
    { 0x0000000c, 0x21043ca5, 0x00000104, 0x00010001 },
    { 0x00000001, 0x21080061, 0x00000000, 0x0001000f },
@@ -72,7 +77,7 @@
    { 0x00600001, 0x40410232, 0x00ae0510, 0x00000000 },
    { 0x00600001, 0x40500232, 0x00ae0521, 0x00000000 },
    { 0x00600001, 0x40510232, 0x00ae0530, 0x00000000 },
-   { 0x01600031, 0x20000c04, 0x508d0000, 0x04082008 },
+   { 0x01600031, 0x20000c04, 0x508d0000, 0x0408200b },
    { 0x01000040, 0x20863dad, 0x00000086, 0xffffffff },
    { 0x00000040, 0x20a03dad, 0x000000a0, 0x00100010 },
    { 0x05000010, 0x200035ac, 0x020000a0, 0x00000084 },
@@ -81,10 +86,10 @@
    { 0x01000010, 0x20003dac, 0x00000086, 0x00010001 },
    { 0x00010001, 0x20b80129, 0x000000c4, 0x00000000 },
    { 0x00010001, 0x20ba0231, 0x000000c6, 0x00000000 },
-   { 0x00010220, 0x34001c00, 0x02001400, 0xffffff64 },
+   { 0x00010220, 0x34001c00, 0x02001400, 0xffffff5a },
    { 0x00000001, 0x20a001ad, 0x0000008a, 0x00000000 },
    { 0x00000040, 0x20a23dad, 0x000000a2, 0x00080008 },
-   { 0x00000220, 0x34001c00, 0x00001400, 0xffffff5e },
+   { 0x00000220, 0x34001c00, 0x00001400, 0xffffff54 },
    { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
    { 0x00600001, 0x21e00022, 0x008d0000, 0x00000000 },
    { 0x0f000031, 0x20000c04, 0x708d0000, 0x82000000 },
diff --git a/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g6b b/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g6b
index 111d483..8d6ebe3 100644
--- a/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g6b
+++ b/src/shaders/post_processing/gen5_6/nv12_dndi_nv12.g6b
@@ -117,18 +117,23 @@
    { 0x00600001, 0x21a00022, 0x008d0100, 0x00000000 },
    { 0x00000001, 0x21c00022, 0x00000560, 0x00000000 },
    { 0x05600031, 0x20000cc4, 0x000001a0, 0x04094014 },
-   { 0x00400040, 0x22083e28, 0x00690024, 0x07000700 },
-   { 0x01000010, 0x20003e2c, 0x0000003b, 0x00010001 },
-   { 0x00010220, 0x34001c00, 0x00001400, 0x0000000a },
-   { 0x00400001, 0x20400022, 0x00690580, 0x00000000 },
-   { 0x00400001, 0x20500022, 0x006904d0, 0x00000000 },
-   { 0x00400001, 0x20600022, 0x00690590, 0x00000000 },
-   { 0x00400001, 0x20700022, 0x006904f0, 0x00000000 },
-   { 0x00000220, 0x34001c00, 0x00001400, 0x00000008 },
-   { 0x00400001, 0x20400022, 0x006904c0, 0x00000000 },
-   { 0x00400001, 0x20500022, 0x00690580, 0x00000000 },
-   { 0x00400001, 0x20600022, 0x006904e0, 0x00000000 },
-   { 0x00400001, 0x20700022, 0x00690590, 0x00000000 },
+   { 0x00800001, 0x20400232, 0x00b10440, 0x00000000 },
+   { 0x00800001, 0x20500232, 0x00b10450, 0x00000000 },
+   { 0x00800001, 0x20600232, 0x00b10460, 0x00000000 },
+   { 0x00800001, 0x20700232, 0x00b10470, 0x00000000 },
+   { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
+   { 0x00000001, 0x21080061, 0x00000000, 0x0003000f },
+   { 0x00600001, 0x20200022, 0x008d0100, 0x00000000 },
+   { 0x05600031, 0x20000cc4, 0x00000020, 0x06094007 },
+   { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
+   { 0x0000000c, 0x21043ca5, 0x00000104, 0x00010001 },
+   { 0x00000001, 0x21080061, 0x00000000, 0x0001000f },
+   { 0x00600001, 0x20200022, 0x008d0100, 0x00000000 },
+   { 0x00600001, 0x40400232, 0x00ae0481, 0x00000000 },
+   { 0x00600001, 0x40410232, 0x00ae0490, 0x00000000 },
+   { 0x00600001, 0x40500232, 0x00ae04a1, 0x00000000 },
+   { 0x00600001, 0x40510232, 0x00ae04b0, 0x00000000 },
+   { 0x05600031, 0x20000cc4, 0x00000020, 0x04094008 },
    { 0x00800001, 0x20400232, 0x00b104c0, 0x00000000 },
    { 0x00800001, 0x20500232, 0x00b104d0, 0x00000000 },
    { 0x00800001, 0x20600232, 0x00b104e0, 0x00000000 },
@@ -136,7 +141,7 @@
    { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
    { 0x00000001, 0x21080061, 0x00000000, 0x0003000f },
    { 0x00600001, 0x20200022, 0x008d0100, 0x00000000 },
-   { 0x05600031, 0x20000cc4, 0x00000020, 0x06094007 },
+   { 0x05600031, 0x20000cc4, 0x00000020, 0x0609400a },
    { 0x00200001, 0x210001a1, 0x004500a0, 0x00000000 },
    { 0x0000000c, 0x21043ca5, 0x00000104, 0x00010001 },
    { 0x00000001, 0x21080061, 0x00000000, 0x0001000f },
@@ -145,7 +150,7 @@
    { 0x00600001, 0x40410232, 0x00ae0510, 0x00000000 },
    { 0x00600001, 0x40500232, 0x00ae0521, 0x00000000 },
    { 0x00600001, 0x40510232, 0x00ae0530, 0x00000000 },
-   { 0x05600031, 0x20000cc4, 0x00000020, 0x04094008 },
+   { 0x05600031, 0x20000cc4, 0x00000020, 0x0409400b },
    { 0x01000040, 0x20863dad, 0x00000086, 0xffffffff },
    { 0x00000040, 0x20a03dad, 0x000000a0, 0x00100010 },
    { 0x05000010, 0x200035ac, 0x020000a0, 0x00000084 },
@@ -154,10 +159,10 @@
    { 0x01000010, 0x20003dac, 0x00000086, 0x00010001 },
    { 0x00010001, 0x20b80129, 0x000000c4, 0x00000000 },
    { 0x00010001, 0x20ba0231, 0x000000c6, 0x00000000 },
-   { 0x00010220, 0x34001c00, 0x02001400, 0xffffff64 },
+   { 0x00010220, 0x34001c00, 0x02001400, 0xffffff5a },
    { 0x00000001, 0x20a001ad, 0x0000008a, 0x00000000 },
    { 0x00000040, 0x20a23dad, 0x000000a2, 0x00080008 },
-   { 0x00000220, 0x34001c00, 0x00001400, 0xffffff5e },
+   { 0x00000220, 0x34001c00, 0x00001400, 0xffffff54 },
    { 0x0000007e, 0x00000000, 0x00000000, 0x00000000 },
    { 0x00600001, 0x21e00022, 0x008d0000, 0x00000000 },
    { 0x07000031, 0x20001cc4, 0x000001e0, 0x82000010 },