[Mesa-dev] Seeking advice speeding up glsl optimisation passes for AoA

[Timothy Arceri]

On Wed, 2015-07-15 at 16:42 +1000, Timothy Arceri wrote:
> Hi guys,
> 
> As I've mentioned a couple of times in previous patches some of the cts AoA
> tests are taking very long time to compile. This is due to excessive
> optimisation passes mainly in the glsl optimisations (there are some 
> slowdowns
> in the intel backend too but these seemed to go away when I tried the new 
> nir
> vec4 backend).
> 
> I fixed part of the problem with this patch to do the dead code elimination 
> in
> a single pass [1]. 
> These excessive passes exist in normal shaders but its generally not an 
> issue
> as the number of passes is generally quite low, and inexpensive. However 
> when
> you have an 8 dimensional array constantly walking this becomes quite
> expensive.
> 
> The remaining issue I'm seeking some advice for is with constant
> propagation/folding.
> 
> It seems for interators used in loops you can get into a situation where an
> optimisation pass is needed for each loop iteration in order to make all
> values of the iterator constant.
> 
> I didn't have look too find some real world examples of this in the public
> shader-db. For example here is it happening for a Unity shader:
> 
> ---------------------- Linker Pass 1 ----------------------
> opt_function_inlining progress
> opt_dead_functions progress
> opt_copy_propagation_elements progress
> opt_tree_grafting progress
> opt_constant_folding progress
> opt_algebraic progress
> ---------------------- Linker Pass 1 ----------------------
> constant prop : l_3, Value : 0
> constant prop : l_3, Value : 0
> constant prop : l_3, Value : 0
> constant prop : l_3, Value : 0
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 2 ----------------------
> constant prop : l_3, Value : 1
> constant prop : l_3, Value : 1
> constant prop : l_3, Value : 1
> constant prop : l_3, Value : 1
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 3 ----------------------
> constant prop : l_3, Value : 2
> constant prop : l_3, Value : 2
> constant prop : l_3, Value : 2
> constant prop : l_3, Value : 2
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 4 ----------------------
> constant prop : l_3, Value : 3
> constant prop : l_3, Value : 3
> constant prop : l_3, Value : 3
> constant prop : l_3, Value : 3
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 5 ----------------------
> constant prop : l_3, Value : 4
> constant prop : l_3, Value : 4
> constant prop : l_3, Value : 4
> constant prop : l_3, Value : 4
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 6 ----------------------
> constant prop : l_3, Value : 5
> constant prop : l_3, Value : 5
> constant prop : l_3, Value : 5
> constant prop : l_3, Value : 5
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 7 ----------------------
> constant prop : l_3, Value : 6
> constant prop : l_3, Value : 6
> constant prop : l_3, Value : 6
> constant prop : l_3, Value : 6
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 8 ----------------------
> constant prop : l_3, Value : 7
> constant prop : l_3, Value : 7
> constant prop : l_3, Value : 7
> constant prop : l_3, Value : 7
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 9 ----------------------
> constant prop : l_3, Value : 8
> constant prop : l_3, Value : 8
> constant prop : l_3, Value : 8
> constant prop : l_3, Value : 8
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 10 ----------------------
> constant prop : l_3, Value : 9
> constant prop : l_3, Value : 9
> constant prop : l_3, Value : 9
> constant prop : l_3, Value : 9
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 11 ----------------------
> constant prop : l_3, Value : 10
> constant prop : l_3, Value : 10
> constant prop : l_3, Value : 10
> constant prop : l_3, Value : 10
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 12 ----------------------
> constant prop : l_3, Value : 11
> constant prop : l_3, Value : 11
> constant prop : l_3, Value : 11
> constant prop : l_3, Value : 11
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 13 ----------------------
> constant prop : l_3, Value : 12
> constant prop : l_3, Value : 12
> constant prop : l_3, Value : 12
> constant prop : l_3, Value : 12
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 14 ----------------------
> constant prop : l_3, Value : 13
> constant prop : l_3, Value : 13
> constant prop : l_3, Value : 13
> constant prop : l_3, Value : 13
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 15 ----------------------
> constant prop : l_3, Value : 14
> constant prop : l_3, Value : 14
> constant prop : l_3, Value : 14
> constant prop : l_3, Value : 14
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 16 ----------------------
> constant prop : l_3, Value : 15
> constant prop : l_3, Value : 15
> constant prop : l_3, Value : 15
> constant prop : l_3, Value : 15
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 17 ----------------------
> constant prop : l_3, Value : 16
> constant prop : l_3, Value : 16
> constant prop : l_3, Value : 16
> constant prop : l_3, Value : 16
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 18 ----------------------
> constant prop : l_3, Value : 17
> constant prop : l_3, Value : 17
> constant prop : l_3, Value : 17
> constant prop : l_3, Value : 17
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 19 ----------------------
> constant prop : l_3, Value : 18
> constant prop : l_3, Value : 18
> constant prop : l_3, Value : 18
> constant prop : l_3, Value : 18
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 20 ----------------------
> constant prop : l_3, Value : 19
> constant prop : l_3, Value : 19
> constant prop : l_3, Value : 19
> constant prop : l_3, Value : 19
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 21 ----------------------
> constant prop : l_3, Value : 20
> constant prop : l_3, Value : 20
> constant prop : l_3, Value : 20
> constant prop : l_3, Value : 20
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 22 ----------------------
> constant prop : l_3, Value : 21
> constant prop : l_3, Value : 21
> constant prop : l_3, Value : 21
> constant prop : l_3, Value : 21
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 23 ----------------------
> constant prop : l_3, Value : 22
> constant prop : l_3, Value : 22
> constant prop : l_3, Value : 22
> constant prop : l_3, Value : 22
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 24 ----------------------
> constant prop : l_3, Value : 23
> constant prop : l_3, Value : 23
> constant prop : l_3, Value : 23
> constant prop : l_3, Value : 23
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 25 ----------------------
> constant prop : l_3, Value : 24
> constant prop : l_3, Value : 24
> constant prop : l_3, Value : 24
> constant prop : l_3, Value : 24
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 26 ----------------------
> constant prop : l_3, Value : 25
> constant prop : l_3, Value : 25
> constant prop : l_3, Value : 25
> constant prop : l_3, Value : 25
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 27 ----------------------
> constant prop : l_3, Value : 26
> constant prop : l_3, Value : 26
> constant prop : l_3, Value : 26
> constant prop : l_3, Value : 26
> opt_constant_propagation progress
> opt_constant_folding progress
> ---------------------- Linker Pass 28 ----------------------
> constant prop : l_3, Value : 27
> constant prop : l_3, Value : 27
> constant prop : l_3, Value : 27
> constant prop : l_3, Value : 27
> opt_constant_propagation progress
> opt_constant_folding progress
> 
> 
> <snip>
> 
> shaders/unity/68.shader_test - FS SIMD8 shader: 1058 inst, 0 loops, 0:0
> spills:fills, Promoted 144 constants, compacted 16928 to 13008 bytes.
> shaders/unity/68.shader_test - FS SIMD16 shader: 1338 inst, 0 loops, 0:0
> spills:fills, Promoted 144 constants, compacted 21408 to 17488 bytes.
> shaders/unity/68.shader_test - VS vec4 shader: 10 inst, 0 loops, compacted 
> 160
> to 144 bytes.
> 
> 
> The problem I'm having is that its expensive to walk over these large arrays
> and in the tests I'm working with the interator goes through 256 
> optimisation
> passes (and then is later dead code eliminated anyway).
> 
> So question for those who have more knowledge about these optimisations then
> me is what would you recommend I do as a next step? Does it make sense to 
> try
> to do more of this work in a single pass or should I just try to avoid
> constantly walking the arrays somehow.
> 
> Thanks for your time,
> Tim
> 
> [1] http://lists.freedesktop.org/archives/mesa-dev/2015-July/088651.html


I used the word iterator throught this email as that was the variable name in
the shader, to be clear I really just meant the loop index/counter.