i965/fs: Lower 32x32 bit multiplication on BXT.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_tex_layout.c
index 9a2a331fcb725e73dc60e4afa3eed04870b17523..fb78b08b64916982f84fff7f67efd5e030a80b00 100644 (file)
@@ -63,7 +63,7 @@ tr_mode_horizontal_texture_alignment(const struct brw_context *brw,
    int i = 0;
 
    /* Alignment computations below assume bpp >= 8 and a power of 2. */
-   assert (bpp >= 8 && bpp <= 128 && is_power_of_two(bpp));
+   assert (bpp >= 8 && bpp <= 128 && _mesa_is_pow_two(bpp));
 
    switch(mt->target) {
    case GL_TEXTURE_1D:
@@ -95,7 +95,7 @@ tr_mode_horizontal_texture_alignment(const struct brw_context *brw,
    ret_align = mt->tr_mode == INTEL_MIPTREE_TRMODE_YF ?
                align_yf[i] : align_ys[i];
 
-   assert(is_power_of_two(mt->num_samples));
+   assert(_mesa_is_pow_two(mt->num_samples));
 
    switch (mt->num_samples) {
    case 2:
@@ -116,8 +116,12 @@ tr_mode_horizontal_texture_alignment(const struct brw_context *brw,
 
 static unsigned int
 intel_horizontal_texture_alignment_unit(struct brw_context *brw,
-                                        struct intel_mipmap_tree *mt)
+                                        struct intel_mipmap_tree *mt,
+                                        uint32_t layout_flags)
 {
+   if (layout_flags & MIPTREE_LAYOUT_FORCE_HALIGN16)
+      return 16;
+
    /**
     * From the "Alignment Unit Size" section of various specs, namely:
     * - Gen3 Spec: "Memory Data Formats" Volume,         Section 1.20.1.4
@@ -172,9 +176,6 @@ intel_horizontal_texture_alignment_unit(struct brw_context *brw,
    if (brw->gen >= 7 && mt->format == MESA_FORMAT_Z_UNORM16)
       return 8;
 
-   if (brw->gen == 8 && mt->mcs_mt && mt->num_samples <= 1)
-      return 16;
-
    return 4;
 }
 
@@ -198,7 +199,7 @@ tr_mode_vertical_texture_alignment(const struct brw_context *brw,
           mt->target != GL_TEXTURE_1D_ARRAY);
 
    /* Alignment computations below assume bpp >= 8 and a power of 2. */
-   assert (bpp >= 8 && bpp <= 128 && is_power_of_two(bpp)) ;
+   assert (bpp >= 8 && bpp <= 128 && _mesa_is_pow_two(bpp)) ;
 
    switch(mt->target) {
    case GL_TEXTURE_2D:
@@ -225,7 +226,7 @@ tr_mode_vertical_texture_alignment(const struct brw_context *brw,
    ret_align = mt->tr_mode == INTEL_MIPTREE_TRMODE_YF ?
                align_yf[i] : align_ys[i];
 
-   assert(is_power_of_two(mt->num_samples));
+   assert(_mesa_is_pow_two(mt->num_samples));
 
    switch (mt->num_samples) {
    case 4:
@@ -365,9 +366,8 @@ brw_miptree_layout_2d(struct intel_mipmap_tree *mt)
 
    mt->total_width = mt->physical_width0;
 
-   if (mt->compressed) {
-       mt->total_width = ALIGN(mt->physical_width0, mt->align_w);
-   }
+   if (mt->compressed)
+       mt->total_width = ALIGN(mt->total_width, bw);
 
    /* May need to adjust width to accommodate the placement of
     * the 2nd mipmap.  This occurs when the alignment
@@ -499,9 +499,9 @@ align_cube(struct intel_mipmap_tree *mt)
       mt->total_height += 2;
 }
 
-static bool
-use_linear_1d_layout(struct brw_context *brw,
-                     struct intel_mipmap_tree *mt)
+bool
+gen9_use_linear_1d_layout(const struct brw_context *brw,
+                          const struct intel_mipmap_tree *mt)
 {
    /* On Gen9+ the mipmap levels of a 1D surface are all laid out in a
     * horizontal line. This isn't done for depth/stencil buffers however
@@ -526,7 +526,7 @@ brw_miptree_layout_texture_array(struct brw_context *brw,
                                 struct intel_mipmap_tree *mt)
 {
    unsigned height = mt->physical_height0;
-   bool layout_1d = use_linear_1d_layout(brw, mt);
+   bool layout_1d = gen9_use_linear_1d_layout(brw, mt);
    int physical_qpitch;
 
    if (layout_1d)
@@ -614,8 +614,8 @@ brw_miptree_layout_texture_3d(struct brw_context *brw,
  */
 static uint32_t
 brw_miptree_choose_tiling(struct brw_context *brw,
-                          enum intel_miptree_tiling_mode requested,
-                          const struct intel_mipmap_tree *mt)
+                          const struct intel_mipmap_tree *mt,
+                          uint32_t layout_flags)
 {
    if (mt->format == MESA_FORMAT_S_UINT8) {
       /* The stencil buffer is W tiled. However, we request from the kernel a
@@ -624,15 +624,18 @@ brw_miptree_choose_tiling(struct brw_context *brw,
       return I915_TILING_NONE;
    }
 
+   /* Do not support changing the tiling for miptrees with pre-allocated BOs. */
+   assert((layout_flags & MIPTREE_LAYOUT_FOR_BO) == 0);
+
    /* Some usages may want only one type of tiling, like depth miptrees (Y
     * tiled), or temporary BOs for uploading data once (linear).
     */
-   switch (requested) {
-   case INTEL_MIPTREE_TILING_ANY:
+   switch (layout_flags & MIPTREE_LAYOUT_ALLOC_ANY_TILED) {
+   case MIPTREE_LAYOUT_ALLOC_ANY_TILED:
       break;
-   case INTEL_MIPTREE_TILING_Y:
+   case MIPTREE_LAYOUT_ALLOC_YTILED:
       return I915_TILING_Y;
-   case INTEL_MIPTREE_TILING_NONE:
+   case MIPTREE_LAYOUT_ALLOC_LINEAR:
       return I915_TILING_NONE;
    }
 
@@ -710,50 +713,10 @@ brw_miptree_choose_tiling(struct brw_context *brw,
    return I915_TILING_Y | I915_TILING_X;
 }
 
-
-void
-brw_miptree_layout(struct brw_context *brw,
-                   bool for_bo,
-                   enum intel_miptree_tiling_mode requested,
-                   struct intel_mipmap_tree *mt)
+static void
+intel_miptree_set_total_width_height(struct brw_context *brw,
+                                     struct intel_mipmap_tree *mt)
 {
-   bool gen6_hiz_or_stencil = false;
-
-   mt->tr_mode = INTEL_MIPTREE_TRMODE_NONE;
-
-   if (brw->gen == 6 && mt->array_layout == ALL_SLICES_AT_EACH_LOD) {
-      const GLenum base_format = _mesa_get_format_base_format(mt->format);
-      gen6_hiz_or_stencil = _mesa_is_depth_or_stencil_format(base_format);
-   }
-
-   if (gen6_hiz_or_stencil) {
-      /* On gen6, we use ALL_SLICES_AT_EACH_LOD for stencil/hiz because the
-       * hardware doesn't support multiple mip levels on stencil/hiz.
-       *
-       * PRM Vol 2, Part 1, 7.5.3 Hierarchical Depth Buffer:
-       * "The hierarchical depth buffer does not support the LOD field"
-       *
-       * PRM Vol 2, Part 1, 7.5.4.1 Separate Stencil Buffer:
-       * "The stencil depth buffer does not support the LOD field"
-       */
-      if (mt->format == MESA_FORMAT_S_UINT8) {
-         /* Stencil uses W tiling, so we force W tiling alignment for the
-          * ALL_SLICES_AT_EACH_LOD miptree layout.
-          */
-         mt->align_w = 64;
-         mt->align_h = 64;
-      } else {
-         /* Depth uses Y tiling, so we force need Y tiling alignment for the
-          * ALL_SLICES_AT_EACH_LOD miptree layout.
-          */
-         mt->align_w = 128 / mt->cpp;
-         mt->align_h = 32;
-      }
-   } else {
-      mt->align_w = intel_horizontal_texture_alignment_unit(brw, mt);
-      mt->align_h = intel_vertical_texture_alignment_unit(brw, mt);
-   }
-
    switch (mt->target) {
    case GL_TEXTURE_CUBE_MAP:
       if (brw->gen == 4) {
@@ -788,7 +751,7 @@ brw_miptree_layout(struct brw_context *brw,
          break;
       case INTEL_MSAA_LAYOUT_NONE:
       case INTEL_MSAA_LAYOUT_IMS:
-         if (use_linear_1d_layout(brw, mt))
+         if (gen9_use_linear_1d_layout(brw, mt))
             gen9_miptree_layout_1d(mt);
          else
             brw_miptree_layout_2d(mt);
@@ -796,8 +759,63 @@ brw_miptree_layout(struct brw_context *brw,
       }
       break;
    }
+
    DBG("%s: %dx%dx%d\n", __func__,
        mt->total_width, mt->total_height, mt->cpp);
+}
+
+static void
+intel_miptree_set_alignment(struct brw_context *brw,
+                            struct intel_mipmap_tree *mt,
+                            uint32_t layout_flags)
+{
+   bool gen6_hiz_or_stencil = false;
+
+   if (brw->gen == 6 && mt->array_layout == ALL_SLICES_AT_EACH_LOD) {
+      const GLenum base_format = _mesa_get_format_base_format(mt->format);
+      gen6_hiz_or_stencil = _mesa_is_depth_or_stencil_format(base_format);
+   }
+
+   if (gen6_hiz_or_stencil) {
+      /* On gen6, we use ALL_SLICES_AT_EACH_LOD for stencil/hiz because the
+       * hardware doesn't support multiple mip levels on stencil/hiz.
+       *
+       * PRM Vol 2, Part 1, 7.5.3 Hierarchical Depth Buffer:
+       * "The hierarchical depth buffer does not support the LOD field"
+       *
+       * PRM Vol 2, Part 1, 7.5.4.1 Separate Stencil Buffer:
+       * "The stencil depth buffer does not support the LOD field"
+       */
+      if (mt->format == MESA_FORMAT_S_UINT8) {
+         /* Stencil uses W tiling, so we force W tiling alignment for the
+          * ALL_SLICES_AT_EACH_LOD miptree layout.
+          */
+         mt->align_w = 64;
+         mt->align_h = 64;
+         assert((layout_flags & MIPTREE_LAYOUT_FORCE_HALIGN16) == 0);
+      } else {
+         /* Depth uses Y tiling, so we force need Y tiling alignment for the
+          * ALL_SLICES_AT_EACH_LOD miptree layout.
+          */
+         mt->align_w = 128 / mt->cpp;
+         mt->align_h = 32;
+      }
+   } else {
+      mt->align_w =
+         intel_horizontal_texture_alignment_unit(brw, mt, layout_flags);
+      mt->align_h = intel_vertical_texture_alignment_unit(brw, mt);
+   }
+}
+
+void
+brw_miptree_layout(struct brw_context *brw,
+                   struct intel_mipmap_tree *mt,
+                   uint32_t layout_flags)
+{
+   mt->tr_mode = INTEL_MIPTREE_TRMODE_NONE;
+
+   intel_miptree_set_alignment(brw, mt, layout_flags);
+   intel_miptree_set_total_width_height(brw, mt);
 
    if (!mt->total_width || !mt->total_height) {
       intel_miptree_release(&mt);
@@ -814,7 +832,7 @@ brw_miptree_layout(struct brw_context *brw,
       mt->align_h /= j;
    }
 
-   if (!for_bo)
-      mt->tiling = brw_miptree_choose_tiling(brw, requested, mt);
+   if ((layout_flags & MIPTREE_LAYOUT_FOR_BO) == 0)
+      mt->tiling = brw_miptree_choose_tiling(brw, mt, layout_flags);
 }