|Summary:||[swrast] piglit glsl-vs-clamp-1 regression|
|Product:||Mesa||Reporter:||Vinson Lee <vlee>|
|Component:||Mesa core||Assignee:||Abdiel Janulgue <abdiel.janulgue>|
|Status:||RESOLVED FIXED||QA Contact:|
|Priority:||medium||CC:||abdiel.janulgue, huax.lu, idr, mattst88|
|i915 platform:||i915 features:|
|Bug Depends on:|
Description Vinson Lee 2014-09-03 23:00:38 UTC
mesa: 58b386dce435d2d82c2dc80b1a8d1373bb3e0ac6 (master 10.4.0-devel) piglit glsl-vs-clamp-1 times out on swrast. 8f890b119eaff88a7fad64abbf183cbcc22edc7a is the first bad commit commit 8f890b119eaff88a7fad64abbf183cbcc22edc7a Author: Abdiel Janulgue <firstname.lastname@example.org> Date: Thu Jun 19 22:15:14 2014 -0700 glsl: Optimize clamp(x, 0, 1) as saturate(x) v2: - Check that the base type is float (Ian Romanick) v3: - Make sure comments reflect that we are doing a commutative operation - Add missing condition where the inner constant is 1.0 and outer constant is 0.0 - Make indexing of operands easier to read (Matt Turner) Reviewed-by: Matt Turner <email@example.com> Reviewed-by: Ian Romanick <firstname.lastname@example.org> Signed-off-by: Abdiel Janulgue <email@example.com> :040000 040000 5174a30266aa05b4334c831e1cd51de1376e4b70 916e9faa0afc6c04e530af2a3c0dfb938a7f07de M src bisect run success
Comment 1 Abdiel Janulgue 2014-09-08 12:57:28 UTC
We got an infinite loop in swrast which is caused by the back and forth ping-pong between saturate-to-clamp lowering pass and the clamp-to-saturate opt_algebraic optimization pass. I got a fix coming up which optionally disables the opt_algebraic optimization by adding an EmitNoSat toggle in gl_shader_compiler_options.
Comment 2 Abdiel Janulgue 2014-09-08 13:18:46 UTC
Created attachment 105892 [details] [review] swrast fix Vinson, can you please try the attached patch if it fixes the issue?
Comment 3 Kenneth Graunke 2014-09-11 05:31:05 UTC
swrast should be able to handle ir_unop_saturate just fine - it shouldn't need the lowering pass at all, even in the vertex shader. This may still be necessary though - I don't know if r200 can do this (or if it even supports GLSL), or about pre-SM3 gallium drivers. It sounds like drivers should set options->EmitNoSat for the VS stages, and we should only add the SAT_TO_CLAMP flag when options->EmitNoSat is set (in addition to disabling the opt_algebraic optimization in that case). That way, saturate actually happens for drivers that can support it.
Comment 4 Vinson Lee 2014-10-11 05:39:06 UTC
attachment 105892 [details] [review] fixes swrast regression. Tested-by: Vinson Lee <firstname.lastname@example.org>
Comment 5 Vinson Lee 2014-10-24 03:55:56 UTC
These other piglit regressions were also introduced with commit 8f890b119eaff88a7fad64abbf183cbcc22edc7a. glsl-vs-raytrace-bug26691 vs-smoothstep-float-float-float vs-smoothstep-float-float-vec2 vs-smoothstep-float-float-vec3 vs-smoothstep-float-float-vec4 vs-smoothstep-vec2-vec2-vec2 vs-smoothstep-vec3-vec3-vec3 vs-smoothstep-vec4-vec4-vec4 fs-bool-less-compare-true vs-saturate-exp2 vs-saturate-pow vs-saturate-sqrt
Comment 6 Abdiel Janulgue 2014-11-28 09:28:57 UTC
(In reply to Kenneth Graunke from comment #3) > swrast should be able to handle ir_unop_saturate just fine - it shouldn't > need the lowering pass at all, even in the vertex shader. > > This may still be necessary though - I don't know if r200 can do this (or if > it even supports GLSL), or about pre-SM3 gallium drivers. > > It sounds like drivers should set options->EmitNoSat for the VS stages, and > we should only add the SAT_TO_CLAMP flag when options->EmitNoSat is set (in > addition to disabling the opt_algebraic optimization in that case). That > way, saturate actually happens for drivers that can support it. Hi Ken! Sorry for the delay. I was not yet sure which drivers you mean. Do you mean *all* drivers (including r200)? Or if I understood this correctly, I would have to set EmitNoSat in the gallium state tracker and in ir_to_mesa as is done in this patch?
Comment 7 Abdiel Janulgue 2014-11-28 09:30:20 UTC
*** Bug 83384 has been marked as a duplicate of this bug. ***
Comment 8 Abdiel Janulgue 2014-12-15 13:45:05 UTC
Patch has now landed in 10.4: http://cgit.freedesktop.org/mesa/mesa/commit/?h=10.4&id=65f03e673310fe1cdf1416d30f14a7f3edc2cd4e