[Mesa-dev] [PATCH 2/5] i965/fs: Emit better b2f of an expression on GEN4 and GEN5

[Ian Romanick]

On 03/16/2015 09:37 AM, Matt Turner wrote:
> On Mon, Mar 16, 2015 at 4:54 AM, Tapani Pälli <tapani.palli at intel.com> wrote:
>> Is there some particular Piglit test case that hits this path and is it
>> possible with gen>5 (by removing gen check)? I've tried this with
>> handicrafted shader_test and also shader-db and cannot hit the conditions
>> for changes to happen. Would be nice to be able to run examine changes and
>> understand this better.
> 
> Well, from the shader-db stats there must be some shaders affected in shader-db.

For example, shaders/anholt/6.shader_test was helped.  That appears to
be the smallest (and one of the only open-source) shaders that is
helped.

helped:   shaders/anholt/6.shader_test FS SIMD16:           86 -> 85 (-1.16%)
helped:   shaders/anholt/6.shader_test FS SIMD8:            77 -> 76 (-1.30%)

The part of that shader that generates the code we care about is:

        gl_FragColor = step(0.0, n_dot_l) *
        vec4((diffuse + vec3(specular)) * shadow, material_color.w);

It should be easy enough to create a test using step() that generates
similar code on GEN4 and GEN5.

> Use INTEL_DEVID_OVERRIDE=... with a Gen4 or Gen5 PCI ID from
> include/pci_ids/i965_pci_ids.h.

That's part of the reason I include the platform name and the PCI ID in
the shader-db log in the commit message. :)