st/xorg: accelerate src luminance

3 files changed, 38 insertions, 21 deletions

diff --git a/src/gallium/state_trackers/xorg/xorg_composite.c b/src/gallium/state_trackers/xorg/xorg_composite.c
index 4dbb490ca5..f16816b2a7 100644
--- a/src/gallium/state_trackers/xorg/xorg_composite.c
+++ b/src/gallium/state_trackers/xorg/xorg_composite.c
@@ -232,15 +232,25 @@ bind_blend_state(struct exa_context *exa, int op,
 }
 
 static unsigned
-picture_format_fixups(struct exa_pixmap_priv *pSrc, PicturePtr pSrcPicture, boolean mask)
+picture_format_fixups(struct exa_pixmap_priv *pSrc, PicturePtr pSrcPicture, boolean mask,
+                      PicturePtr pDstPicture)
 {
    boolean set_alpha = FALSE;
    boolean swizzle = FALSE;
    unsigned ret = 0;
 
    if (pSrc->picture_format == pSrcPicture->format) {
-      if (pSrc->picture_format == PICT_a8)
-         return mask ? FS_MASK_LUMINANCE : FS_SRC_LUMINANCE;
+      if (pSrc->picture_format == PICT_a8) {
+         if (mask)
+            return FS_MASK_LUMINANCE;
+         else if (pDstPicture->format != PICT_a8) {
+            /* if both dst and src are luminance then
+             * we don't want to swizzle the alpha (X) of the
+             * source into W component of the dst because
+             * it will break our destination */
+            return FS_SRC_LUMINANCE;
+         }
+      }
       return 0;
    }
 
@@ -285,7 +295,7 @@ picture_format_fixups(struct exa_pixmap_priv *pSrc, PicturePtr pSrcPicture, bool
 
 static void
 bind_shaders(struct exa_context *exa, int op,
-             PicturePtr pSrcPicture, PicturePtr pMaskPicture,
+             PicturePtr pSrcPicture, PicturePtr pMaskPicture, PicturePtr pDstPicture,
              struct exa_pixmap_priv *pSrc, struct exa_pixmap_priv *pMask)
 {
    unsigned vs_traits = 0, fs_traits = 0;
@@ -313,7 +323,7 @@ bind_shaders(struct exa_context *exa, int op,
          vs_traits |= VS_COMPOSITE;
       }
 
-      fs_traits |= picture_format_fixups(pSrc, pSrcPicture, FALSE);
+      fs_traits |= picture_format_fixups(pSrc, pSrcPicture, FALSE, pDstPicture);
    }
 
    if (pMaskPicture) {
@@ -331,7 +341,7 @@ bind_shaders(struct exa_context *exa, int op,
             fs_traits |= FS_CA_FULL;
       }
 
-      fs_traits |= picture_format_fixups(pMask, pMaskPicture, TRUE);
+      fs_traits |= picture_format_fixups(pMask, pMaskPicture, TRUE, pDstPicture);
    }
 
    shader = xorg_shaders_get(exa->renderer->shaders, vs_traits, fs_traits);
@@ -497,7 +507,7 @@ boolean xorg_composite_bind_state(struct exa_context *exa,
    renderer_bind_viewport(exa->renderer, pDst);
    bind_blend_state(exa, op, pSrcPicture, pMaskPicture, pDstPicture);
    renderer_bind_rasterizer(exa->renderer);
-   bind_shaders(exa, op, pSrcPicture, pMaskPicture, pSrc, pMask);
+   bind_shaders(exa, op, pSrcPicture, pMaskPicture, pDstPicture, pSrc, pMask);
    bind_samplers(exa, op, pSrcPicture, pMaskPicture,
                  pDstPicture, pSrc, pMask, pDst);
    setup_constant_buffers(exa, pDst);
diff --git a/src/gallium/state_trackers/xorg/xorg_exa.c b/src/gallium/state_trackers/xorg/xorg_exa.c
index 3e2f4e8834..aa04586455 100644
--- a/src/gallium/state_trackers/xorg/xorg_exa.c
+++ b/src/gallium/state_trackers/xorg/xorg_exa.c
@@ -578,8 +578,6 @@ ExaPrepareComposite(int op, PicturePtr pSrcPicture,
                        render_format_name(priv->picture_format),
                        render_format_name(pSrcPicture->format));
 
-      if (priv->picture_format == PICT_a8)
-         XORG_FALLBACK("pSrc pic_format == PICT_a8");
    }
 
    if (pMask) {
diff --git a/src/gallium/state_trackers/xorg/xorg_exa_tgsi.c b/src/gallium/state_trackers/xorg/xorg_exa_tgsi.c
index 3bf64b6331..13a9840bdd 100644
--- a/src/gallium/state_trackers/xorg/xorg_exa_tgsi.c
+++ b/src/gallium/state_trackers/xorg/xorg_exa_tgsi.c
@@ -396,15 +396,11 @@ xrender_tex(struct ureg_program *ureg,
             struct ureg_dst dst,
             struct ureg_src coords,
             struct ureg_src sampler,
+            struct ureg_src imm0,
             boolean repeat_none,
             boolean swizzle,
             boolean set_alpha)
 {
-   struct ureg_src imm0 = { 0 };
-
-   if (repeat_none || set_alpha)
-      imm0 = ureg_imm4f(ureg, 0, 0, 0, 1);
-
    if (repeat_none) {
       struct ureg_dst tmp0 = ureg_DECL_temporary(ureg);
       struct ureg_dst tmp1 = ureg_DECL_temporary(ureg);
@@ -466,6 +462,7 @@ create_fs(struct pipe_context *pipe,
    struct ureg_src /*dst_pos,*/ src_input, mask_pos;
    struct ureg_dst src, mask;
    struct ureg_dst out;
+   struct ureg_src imm0 = { 0 };
    unsigned has_mask = (fs_traits & FS_MASK) != 0;
    unsigned is_fill = (fs_traits & FS_FILL) != 0;
    unsigned is_composite = (fs_traits & FS_COMPOSITE) != 0;
@@ -483,8 +480,6 @@ create_fs(struct pipe_context *pipe,
    unsigned src_luminance = (fs_traits & FS_SRC_LUMINANCE) != 0;
    unsigned mask_luminance = (fs_traits & FS_MASK_LUMINANCE) != 0;
 
-   if (src_luminance)
-      assert(!"src_luminance not supported");
 #if 0
    print_fs_traits(fs_traits);
 #else
@@ -502,6 +497,11 @@ create_fs(struct pipe_context *pipe,
                           TGSI_SEMANTIC_COLOR,
                           0);
 
+   if (src_repeat_none || mask_repeat_none ||
+       src_set_alpha || mask_set_alpha ||
+       src_luminance) {
+      imm0 = ureg_imm4f(ureg, 0, 0, 0, 1);
+   }
    if (is_composite) {
       src_sampler = ureg_DECL_sampler(ureg, 0);
       src_input = ureg_DECL_fs_input(ureg,
@@ -540,16 +540,17 @@ create_fs(struct pipe_context *pipe,
                                 TGSI_INTERPOLATE_PERSPECTIVE);
 #endif
 
+
    if (is_composite) {
-      if (has_mask)
+      if (has_mask || src_luminance)
          src = ureg_DECL_temporary(ureg);
       else
          src = out;
-      xrender_tex(ureg, src, src_input, src_sampler,
+      xrender_tex(ureg, src, src_input, src_sampler, imm0,
                   src_repeat_none, src_swizzle, src_set_alpha);
    } else if (is_fill) {
       if (is_solid) {
-         if (has_mask)
+         if (has_mask || src_luminance)
             src = ureg_dst(src_input);
          else
             ureg_MOV(ureg, out, src_input);
@@ -557,7 +558,7 @@ create_fs(struct pipe_context *pipe,
          struct ureg_src coords, const0124,
             matrow0, matrow1, matrow2;
 
-         if (has_mask)
+         if (has_mask || src_luminance)
             src = ureg_DECL_temporary(ureg);
          else
             src = out;
@@ -582,10 +583,18 @@ create_fs(struct pipe_context *pipe,
       } else
          debug_assert(!"Unknown fill type!");
    }
+   if (src_luminance) {
+      ureg_MOV(ureg, src,
+               ureg_scalar(ureg_src(src), TGSI_SWIZZLE_X));
+      ureg_MOV(ureg, ureg_writemask(src, TGSI_WRITEMASK_XYZ),
+               ureg_scalar(imm0, TGSI_SWIZZLE_X));
+      if (!has_mask)
+         ureg_MOV(ureg, out, ureg_src(src));
+   }
 
    if (has_mask) {
       mask = ureg_DECL_temporary(ureg);
-      xrender_tex(ureg, mask, mask_pos, mask_sampler,
+      xrender_tex(ureg, mask, mask_pos, mask_sampler, imm0,
                   mask_repeat_none, mask_swizzle, mask_set_alpha);
       /* src IN mask */
       src_in_mask(ureg, out, ureg_src(src), ureg_src(mask),