[Mesa-dev] [PATCH 7/9] gallium/u_tests: test a NULL texture sampler view

[Marek Olšák]

Interesting. On radeonsi, the default value of an unbound texture can
be chosen but is the same for "load", "sample", and "gather"
instructions. We can't return (0,0,0,0) for "load" and (0,0,0,1) for
"sample".

I'll just modify the test to accept both return values.

Marek

On Mon, Feb 2, 2015 at 5:19 PM, Roland Scheidegger <sroland at vmware.com> wrote:
> Actually, since ARB_robust_buffer_access_behavior, GL requires some
> specfic values, but it's - odd...
> "In all the above cases, if the context was created with robust buffer
> access enabled then instead of undefined behavior the result of the
> texel fetch is zero. For the case of a texel fetch on an incomplete
> texture the result is a texture source color of (0, 0, 0, 1)."
> So, it seems if the texture was incomplete, then 0/0/0/1 has to be
> returned, but if it was unbound, then all zeros is the right answer.
> However, this is only true for textures _fetches_, not ordinary sampling
> (it is described in the vertex shader functionality). I would assume
> this to be the case for ordinary sampling too, but there the spec only
> describes the 0/0/0/1 behavior for incomplete textures, not the behavior
> for unbound ones...
>
> Roland
>
>
>
> Am 02.02.2015 um 16:51 schrieb Roland Scheidegger:
>> I don't think this is really correct. llvmpipe will return all zeros on
>> purpose, because this is d3d10 behavior (and dummy textures cannot work
>> correctly with d3d10). Traditionally both d3d9 and gl state trackers
>> used dummy textures, though I'm unsure what values they required (if
>> any). (For d3d9 though 0/0/0/1 would be somewhat odd since it would be
>> different to what bound textures but unused channels have to return, but
>> d3d9 is sometimes odd so it's possible. Well in that way d3d10 is odd
>> too...)
>> I certainly do agree though with the "don't crash" part.
>>
>> Am 01.02.2015 um 18:15 schrieb Marek Olšák:
>>> From: Marek Olšák <marek.olsak at amd.com>
>>>
>>> It shouldn't crash the GPU and it should return (0, 0, 0, 1).
>>>
>>> This is r300 behavior, so I assume it's also DX9 behavior.
>>> Radeonsi can support this easily.
>>> ---
>>>  src/gallium/auxiliary/util/u_tests.c | 39 ++++++++++++++++++++++++++++++++++++
>>>  1 file changed, 39 insertions(+)
>>>
>>> diff --git a/src/gallium/auxiliary/util/u_tests.c b/src/gallium/auxiliary/util/u_tests.c
>>> index 89ae840..9c8514c 100644
>>> --- a/src/gallium/auxiliary/util/u_tests.c
>>> +++ b/src/gallium/auxiliary/util/u_tests.c
>>> @@ -304,6 +304,44 @@ tgsi_vs_window_space_position(struct pipe_context *ctx)
>>>     util_report_result(pass);
>>>  }
>>>
>>> +static void
>>> +null_sampler_view(struct pipe_context *ctx)
>>> +{
>>> +   struct cso_context *cso;
>>> +   struct pipe_resource *cb;
>>> +   void *fs, *vs;
>>> +   bool pass = true;
>>> +   static const float expected[] = {0, 0, 0, 1};
>>> +
>>> +   cso = cso_create_context(ctx);
>>> +   cb = util_create_texture2d(ctx->screen, 256, 256,
>>> +                              PIPE_FORMAT_R8G8B8A8_UNORM);
>>> +   util_set_common_states_and_clear(cso, ctx, cb);
>>> +
>>> +   ctx->set_sampler_views(ctx, PIPE_SHADER_FRAGMENT, 0, 1, NULL);
>>> +
>>> +   /* Fragment shader. */
>>> +   fs = util_make_fragment_tex_shader(ctx, TGSI_TEXTURE_2D,
>>> +                                      TGSI_INTERPOLATE_LINEAR);
>>> +   cso_set_fragment_shader_handle(cso, fs);
>>> +
>>> +   /* Vertex shader. */
>>> +   vs = util_set_passthrough_vertex_shader(cso, ctx, false);
>>> +   util_draw_fullscreen_quad(cso);
>>> +
>>> +   /* Probe pixels. */
>>> +   pass = pass && util_probe_rect_rgba(ctx, cb, 0, 0,
>>> +                                       cb->width0, cb->height0, expected);
>>> +
>>> +   /* Cleanup. */
>>> +   cso_destroy_context(cso);
>>> +   ctx->delete_vs_state(ctx, vs);
>>> +   ctx->delete_fs_state(ctx, fs);
>>> +   pipe_resource_reference(&cb, NULL);
>>> +
>>> +   util_report_result(pass);
>>> +}
>>> +
>>>  /**
>>>   * Run all tests. This should be run with a clean context after
>>>   * context_create.
>>> @@ -314,6 +352,7 @@ util_run_tests(struct pipe_screen *screen)
>>>     struct pipe_context *ctx = screen->context_create(screen, NULL);
>>>
>>>     tgsi_vs_window_space_position(ctx);
>>> +   null_sampler_view(ctx);
>>>
>>>     ctx->destroy(ctx);
>>>
>>>
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.freedesktop.org_mailman_listinfo_mesa-2Ddev&d=AwIGaQ&c=Sqcl0Ez6M0X8aeM67LKIiDJAXVeAw-YihVMNtXt-uEs&r=Vjtt0vs_iqoI31UfJxBl7yv9I2FeiaeAYgMTLKRBc_I&m=XOQSHQnIu1sXl65rM4KSLIO5-Zej5H2L9o1CVQNkXvQ&s=uBbprkG3Qul-bWCqNmg5fX-4b9csQamQpx5wWJZmWcU&e=
>>
>