fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-07 02:48:06 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

glsl: Don't copy propagate elements from SSBO or shared variables either

Browse source 
Since SSBOs can be written by a different GPU thread, copy propagating a
read can cause the value to magically change.  SSBO reads are also very
expensive, so doing it twice will be slower.

The same shader was helped by this patch and the previous.

Haswell, Broadwell, and Skylake had similar results. (Skylake shown)
total instructions in shared programs: 14399119 -> 14399113 (<.01%)
instructions in affected programs: 683 -> 677 (-0.88%)
helped: 1
HURT: 0

total cycles in shared programs: 532973113 -> 532971865 (<.01%)
cycles in affected programs: 524666 -> 523418 (-0.24%)
helped: 1
HURT: 0

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Cc: mesa-stable@lists.freedesktop.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=106774
(cherry picked from commit 37bd9ccd21)
This commit is contained in:
Ian Romanick 2018-06-05 16:02:25 -07:00 • committed by Dylan Baker
parent f3ec346ab1
commit a4609fe84f
1 changed files with 8 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
src/compiler/glsl/opt_copy_propagation_elements.cpp
View file

@ -544,6 +544,10 @@ ir_copy_propagation_elements_visitor::add_copy(ir_assignment *ir)
if (!lhs || !(lhs->type->is_scalar() || lhs->type->is_vector()))
return;
if (lhs->var->data.mode == ir_var_shader_storage ||
lhs->var->data.mode == ir_var_shader_shared)
return;
ir_dereference_variable *rhs = ir->rhs->as_dereference_variable();
if (!rhs) {
ir_swizzle *swiz = ir->rhs->as_swizzle();
@ -560,6 +564,10 @@ ir_copy_propagation_elements_visitor::add_copy(ir_assignment *ir)
orig_swizzle[3] = swiz->mask.w;
}
if (rhs->var->data.mode == ir_var_shader_storage ||
rhs->var->data.mode == ir_var_shader_shared)
return;
/* Move the swizzle channels out to the positions they match in the
* destination. We don't want to have to rewrite the swizzle[]
* array every time we clear a bit of the write_mask.