i965: compute DDX in a subspan based only on top row

Consider only the top-left and top-right pixels to approximate DDX in a 2x2
subspan, unless the application requests a more accurate approximation via
GL_FRAGMENT_SHADER_DERIVATIVE_HINT or this optimization is disabled from the
new driconf option disable_derivative_optimization.

This results in a less accurate approximation.  However, it improves the
performance of Xonotic with Ultra settings by 24.3879% +/- 0.832202% (at 95.0%
confidence) on Haswell.  No noticeable image quality difference observed.

The improvement comes from faster sample_d.  It seems, on Haswell, some
optimizations are introduced to allow faster sample_d when all pixels in a
subspan have the same derivative.  I considered SAMPLE_STATE too, which allows
one to control the quality of sample_d on Haswell.  But it gave much worse
image quality without giving better performance comparing to this change.

No piglit quick.tests regression on Haswell (tested with v1).

v2: better guess for precompile program key

Signed-off-by: Chia-I Wu <olv@lunarg.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Chris Forbes <chrisf@ijw.co.nz>

diff --git a/src/mesa/drivers/dri/i965/brw_context.c b/src/mesa/drivers/dri/i965/brw_context.c
index 5f58a291d8d158fa579616a28fa6543ca7cd6114..18b8e573b598518d0abf383363cab952b6911171 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_context.c
+++ b/src/mesa/drivers/dri/i965/brw_context.c
@@ -478,6 +478,8 @@ brwCreateContext(int api,
    brw_draw_init( brw );
 
    brw->precompile = driQueryOptionb(&brw->optionCache, "shader_precompile");
+   brw->disable_derivative_optimization =
+      driQueryOptionb(&brw->optionCache, "disable_derivative_optimization");
 
    ctx->Const.ContextFlags = 0;
    if ((flags & __DRI_CTX_FLAG_FORWARD_COMPATIBLE) != 0)

diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index 0f88bad247565f45864a797ff6f609b36f65398a..0ec12185ee4978af1fa1c68afd246f5f4bebb0b9 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -1005,6 +1005,7 @@ struct brw_context
    bool always_flush_cache;
    bool disable_throttling;
    bool precompile;
+   bool disable_derivative_optimization;
 
    driOptionCache optionCache;
    /** @} */

diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index bcb15ee0f92323420c63eb8118f6e67c6b5ad2fc..0b441d451f15b8bdd1cf582ac56f32c40abb16d8 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -3213,6 +3213,12 @@ brw_fs_precompile(struct gl_context *ctx, struct gl_shader_program *prog)
 
    key.nr_color_regions = 1;
 
+   /* GL_FRAGMENT_SHADER_DERIVATIVE_HINT is almost always GL_DONT_CARE.  The
+    * quality of the derivatives is likely to be determined by the driconf
+    * option.
+    */
+   key.high_quality_derivatives = brw->disable_derivative_optimization;
+
    key.program_string_id = bfp->id;
 
    uint32_t old_prog_offset = brw->wm.base.prog_offset;

diff --git a/src/mesa/drivers/dri/i965/brw_fs_generator.cpp b/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
index 7ce42c4b9dce5970d3af727d61b27b13f9962092..9eb5e177928ef81d8c4f240e9a329b5ca6973a05 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_generator.cpp
@@ -540,7 +540,7 @@ fs_generator::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src
  *
  * arg0: ss0.tl ss0.tr ss0.bl ss0.br ss1.tl ss1.tr ss1.bl ss1.br
  *
- * and we're trying to produce:
+ * Ideally, we want to produce:
  *
  *           DDX                     DDY
  * dst: (ss0.tr - ss0.tl)     (ss0.tl - ss0.bl)
@@ -556,24 +556,41 @@ fs_generator::generate_tex(fs_inst *inst, struct brw_reg dst, struct brw_reg src
  *
  * For DDX, it ends up being easy: width = 2, horiz=0 gets us the same result
  * for each pair, and vertstride = 2 jumps us 2 elements after processing a
- * pair. But for DDY, it's harder, as we want to produce the pairs swizzled
- * between each other.  We could probably do it like ddx and swizzle the right
- * order later, but bail for now and just produce
+ * pair.  But the ideal approximation may impose a huge performance cost on
+ * sample_d.  On at least Haswell, sample_d instruction does some
+ * optimizations if the same LOD is used for all pixels in the subspan.
+ *
+ * For DDY, it's harder, as we want to produce the pairs swizzled between each
+ * other.  We could probably do it like ddx and swizzle the right order later,
+ * but bail for now and just produce
  * ((ss0.tl - ss0.bl)x4 (ss1.tl - ss1.bl)x4)
  */
 void
 fs_generator::generate_ddx(fs_inst *inst, struct brw_reg dst, struct brw_reg src)
 {
+   unsigned vstride, width;
+
+   if (c->key.high_quality_derivatives) {
+      /* produce accurate derivatives */
+      vstride = BRW_VERTICAL_STRIDE_2;
+      width = BRW_WIDTH_2;
+   }
+   else {
+      /* replicate the derivative at the top-left pixel to other pixels */
+      vstride = BRW_VERTICAL_STRIDE_4;
+      width = BRW_WIDTH_4;
+   }
+
    struct brw_reg src0 = brw_reg(src.file, src.nr, 1,
                                 BRW_REGISTER_TYPE_F,
-                                BRW_VERTICAL_STRIDE_2,
-                                BRW_WIDTH_2,
+                                vstride,
+                                width,
                                 BRW_HORIZONTAL_STRIDE_0,
                                 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
    struct brw_reg src1 = brw_reg(src.file, src.nr, 0,
                                 BRW_REGISTER_TYPE_F,
-                                BRW_VERTICAL_STRIDE_2,
-                                BRW_WIDTH_2,
+                                vstride,
+                                width,
                                 BRW_HORIZONTAL_STRIDE_0,
                                 BRW_SWIZZLE_XYZW, WRITEMASK_XYZW);
    brw_ADD(p, dst, src0, negate(src1));

diff --git a/src/mesa/drivers/dri/i965/brw_wm.c b/src/mesa/drivers/dri/i965/brw_wm.c
index 3d7ca2a4d038b096461843838fdaef16526ca9a5..9745edac6cae0f353e110b6777c63acb9c8cb19b 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_wm.c
+++ b/src/mesa/drivers/dri/i965/brw_wm.c
@@ -416,6 +416,15 @@ static void brw_wm_populate_key( struct brw_context *brw,
 
    key->line_aa = line_aa;
 
+   /* _NEW_HINT */
+   if (brw->disable_derivative_optimization) {
+      key->high_quality_derivatives =
+         ctx->Hint.FragmentShaderDerivative != GL_FASTEST;
+   } else {
+      key->high_quality_derivatives =
+         ctx->Hint.FragmentShaderDerivative == GL_NICEST;
+   }
+
    if (brw->gen < 6)
       key->stats_wm = brw->stats_wm;
 
@@ -503,6 +512,7 @@ const struct brw_tracked_state brw_wm_prog = {
                _NEW_STENCIL |
                _NEW_POLYGON |
                _NEW_LINE |
+               _NEW_HINT |
                _NEW_LIGHT |
                _NEW_FRAG_CLAMP |
                _NEW_BUFFERS |

diff --git a/src/mesa/drivers/dri/i965/brw_wm.h b/src/mesa/drivers/dri/i965/brw_wm.h
index f7a2c5f234da5aaa78314cac1b8c4ae4eed3f4a1..aa786def463963d2d449409c01a8a89b47c94876 100644 (file)
--- a/src/mesa/drivers/dri/i965/brw_wm.h
+++ b/src/mesa/drivers/dri/i965/brw_wm.h
@@ -66,6 +66,7 @@ struct brw_wm_prog_key {
    GLuint render_to_fbo:1;
    GLuint clamp_fragment_color:1;
    GLuint line_aa:2;
+   GLuint high_quality_derivatives:1;
 
    GLushort drawable_height;
    GLbitfield64 input_slots_valid;

diff --git a/src/mesa/drivers/dri/i965/intel_screen.c b/src/mesa/drivers/dri/i965/intel_screen.c
index de80a00ebb35cc054195c744e9b1b1158430c972..cddc8e8133079cbbdeb70d94770b9fb1fe64bc40 100644 (file)
--- a/src/mesa/drivers/dri/i965/intel_screen.c
+++ b/src/mesa/drivers/dri/i965/intel_screen.c
@@ -58,6 +58,10 @@ PUBLIC const char __driConfigOptions[] =
         DRI_CONF_DESC(en, "Enable Hierarchical Z on gen6+")
       DRI_CONF_OPT_END
 
+      DRI_CONF_OPT_BEGIN_B(disable_derivative_optimization, "false")
+        DRI_CONF_DESC(en, "Derivatives with finer granularity by default")
+      DRI_CONF_OPT_END
+
    DRI_CONF_SECTION_END
    DRI_CONF_SECTION_QUALITY
       DRI_CONF_FORCE_S3TC_ENABLE("false")