[Mesa-dev] [PATCH 07/22] i965/meta: Store source scale and offset as vec2s

[Ian Romanick]

From: Ian Romanick <ian.d.romanick at intel.com>

This has a few small benefits.

1. Reduces the size (and the tracking size) of the shaders.  This should
basically undo the text expansion done previously.

2. Reduces the number of _mesa_Uniform* calls from 4 to 2.  This means
half as much time spent doing redundant validation of parameters, etc.

3. Allows the use of _mesa_Uniform*v.  On x64, the non-v path takes the
value in a register, stores it on the stack, and calls _mesa_uniform
with a pointer to that stack location.  A future patch could call
_mesa_uniform directly.

This patch and the next couple could probably be squashed.  I authored
them in several steps to make debugging (bisecting) easier and to make
reviewing easier.

   text	   data	    bss	    dec	    hex	filename
5126824	 209888	  28120	5364832	 51dc60	before-64/lib64/i965_dri.so
5126696	 209888	  28120	5364704	 51dbe0	after-64/lib64/i965_dri.so

Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
---
 src/mesa/drivers/dri/i965/brw_meta_stencil_blit.c | 54 ++++++++++++-----------
 1 file changed, 28 insertions(+), 26 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_meta_stencil_blit.c b/src/mesa/drivers/dri/i965/brw_meta_stencil_blit.c
index a9fb43a..a48034e 100644
--- a/src/mesa/drivers/dri/i965/brw_meta_stencil_blit.c
+++ b/src/mesa/drivers/dri/i965/brw_meta_stencil_blit.c
@@ -101,10 +101,8 @@ static const struct sampler_and_fetch {
 /* We can't use an enum for this because of the preprocessor stringify
  * techinique.
  */
-#define src_x_scale_loc 0
-#define src_y_scale_loc 1
-#define src_x_off_loc 2
-#define src_y_off_loc 3
+#define src_scale_loc 0
+#define src_off_loc 1
 #define dst_x_off_loc 4
 #define dst_y_off_loc 5
 #define draw_rect_w_loc 6
@@ -130,12 +128,10 @@ static const char *fs_tmpl =
    "#extension GL_ARB_explicit_attrib_location: require\n"
    "#extension GL_ARB_explicit_uniform_location: require\n"
    "%s"
-   DECLARE_UNIFORM(float, src_x_scale)
-   DECLARE_UNIFORM(float, src_y_scale)
+   DECLARE_UNIFORM(vec2, src_scale)
 
    /* Top right coordinates of the source rectangle in W-tiled space. */
-   DECLARE_UNIFORM(float, src_x_off)
-   DECLARE_UNIFORM(float, src_y_off)
+   DECLARE_UNIFORM(vec2, src_off)
 
    /* Top right coordinates of the target rectangle in Y-tiled space. */
    DECLARE_UNIFORM(float, dst_x_off)
@@ -167,8 +163,8 @@ static const char *fs_tmpl =
    "\n"
    "void translate_dst_to_src()\n"
    "{\n"
-   "   txl_coords.x = int(float(txl_coords.x) * src_x_scale + src_x_off);\n"
-   "   txl_coords.y = int(float(txl_coords.y) * src_y_scale + src_y_off);\n"
+   "   txl_coords.x = int(float(txl_coords.x) * src_scale.x + src_off.x);\n"
+   "   txl_coords.y = int(float(txl_coords.y) * src_scale.y + src_off.y);\n"
    "}\n"
    "\n"
    "void translate_y_to_w_tiling()\n"
@@ -278,17 +274,17 @@ setup_drawing_rect(const struct blit_dims *dims)
  *   src_x = src_x0 + (dst_x1 -dst_x - 0.5) * scale
  */
 static void
-setup_coord_coeff(GLuint multiplier, GLuint offset,
-                  int src_0, int src_1, int dst_0, int dst_1, bool mirror)
+calculate_coord_coeff(float *scale, float *offset,
+                      int src_0, int src_1, int dst_0, int dst_1, bool mirror)
 {
-   const float scale = ((float)(src_1 - src_0)) / (dst_1 - dst_0);
+   const float s = ((float)(src_1 - src_0)) / (dst_1 - dst_0);
 
    if (mirror) {
-      _mesa_Uniform1f(multiplier, -scale);
-      _mesa_Uniform1f(offset, src_0 + (dst_1 - 0.5f) * scale);
+      *scale = -s;
+      *offset = src_0 + (dst_1 - 0.5f) * s;
    } else {
-      _mesa_Uniform1f(multiplier, scale);
-      _mesa_Uniform1f(offset, src_0 + (-dst_0 + 0.5f) * scale);
+      *scale = s;
+      *offset = src_0 + (-dst_0 + 0.5f) * s;
    }
 }
 
@@ -304,15 +300,21 @@ setup_coord_coeff(GLuint multiplier, GLuint offset,
 static void
 setup_coord_transform(const struct blit_dims *dims)
 {
-   setup_coord_coeff(src_x_scale_loc,
-                     src_x_off_loc,
-                     dims->src_x0, dims->src_x1, dims->dst_x0, dims->dst_x1,
-                     dims->mirror_x);
-
-   setup_coord_coeff(src_y_scale_loc,
-                     src_y_off_loc,
-                     dims->src_y0, dims->src_y1, dims->dst_y0, dims->dst_y1,
-                     dims->mirror_y);
+   float scale[2];
+   float offset[2];
+
+   calculate_coord_coeff(&scale[0],
+                         &offset[0],
+                         dims->src_x0, dims->src_x1, dims->dst_x0, dims->dst_x1,
+                         dims->mirror_x);
+
+   calculate_coord_coeff(&scale[1],
+                         &offset[1],
+                         dims->src_y0, dims->src_y1, dims->dst_y0, dims->dst_y1,
+                         dims->mirror_y);
+
+   _mesa_Uniform2fv(src_scale_loc, 1, scale);
+   _mesa_Uniform2fv(src_off_loc, 1, offset);
 }
 
 static void
-- 
2.5.0