llvmpipe: use simple coeffs calc for 128bit vectors

There are currently two methods in llvmpipe code to calculate coeffs to
be used as inputs for the fragment shader. The two methods use slightly
different ways to do the floating point calculations and thus produce
slightly different results.

The decision which method to use is determined by the size of the vector
that is used by the platform.

For vectors with size of more than 128bit, a single-step method is used,
in which coeffs_init_simple() + attribs_update_simple() are called.

For vectors with size of 128bit or less, a two-step method is used, in
which coeffs_init() + attribs_update() are called.

This causes some piglit tests (clip-distance-bulk-copy,
interface-vs-unnamed-to-fs-unnamed) to fail when using platforms with
128bit vectors (such as ppc64le or x86-64 without AVX).

This patch makes platforms with 128bit vectors use the single-step
method (aka "simple" method) instead of the two-step method.
This would make the resulting coeffs identical between more platforms,
make sure the piglit tests passes, and make debugging and maintainability
a bit easier as the generated LLVM IR will be the same for more platforms.

The performance impact is negligible for x86-64 without AVX, and
basically non-existent for ppc64le, as it can be seen from the following
benchmarking results:

- glxspheres, on ppc64le:

   - original code:  4.892745317 frames/sec 5.460303857 Mpixels/sec
   - with the patch: 4.932083873 frames/sec 5.504205571 Mpixels/sec
   - Additional 0.8% performance boost

- glxspheres, on x86-64 without AVX:

   - original code:  20.16418809 frames/sec 22.50323395 Mpixels/sec
   - with the patch: 20.31328989 frames/sec 22.66963152 Mpixels/sec
   - Additional 0.74% performance boost

- glmark2, on ppc64le:

  - original code:  score of 58
  - with my change: score of 57

- glmark2, on x86-64 without AVX:

  - original code:  score of 175
  - with the patch: score of 167
  - Impact of of -4.5% on performance

- OpenArena, on ppc64le:

  - original code:  3398 frames 1719.0 seconds 2.0 fps
                    255.0/505.9/2773.0/0.0 ms

  - with the patch: 3398 frames 1690.4 seconds 2.0 fps
                    241.0/497.5/2563.0/0.2 ms

  - 29 seconds faster with the patch, which is about 2%

- OpenArena, on x86-64 without AVX:

  - original code:  3398 frames 239.6 seconds 14.2 fps
                    38.0/70.5/719.0/14.6 ms

  - with the patch: 3398 frames 244.4 seconds 13.9 fps
                    38.0/71.9/697.0/14.3 ms

  - 0.3 fps slower with the patch (about 2%)

Additional details can be found at:
http://lists.freedesktop.org/archives/mesa-dev/2015-October/098635.html

Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>

1 files changed, 6 insertions, 1 deletions

diff --git a/src/gallium/drivers/llvmpipe/lp_bld_interp.c b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
index df262fa4716..ceac86abe1d 100644
--- a/src/gallium/drivers/llvmpipe/lp_bld_interp.c
+++ b/src/gallium/drivers/llvmpipe/lp_bld_interp.c
@@ -746,7 +746,12 @@ lp_build_interp_soa_init(struct lp_build_interp_soa_context *bld,
 
    pos_init(bld, x0, y0);
 
-   if (coeff_type.length > 4) {
+   /*
+    * Simple method (single step interpolation) may be slower if vector length
+    * is just 4, but the results are different (generally less accurate) with
+    * the other method, so always use more accurate version.
+    */
+   if (1) {
       bld->simple_interp = TRUE;
       {
          /* XXX this should use a global static table */