<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<p><br>
</p>
<div class="moz-cite-prefix">On 4/23/19 5:16 PM, Jason Ekstrand
wrote:<br>
</div>
<blockquote type="cite"
cite="mid:CAOFGe94RBM09CMhhYpe0cgTpdQdRDN3y2VmGwFExpXpWhuvW8w@mail.gmail.com">
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
<div dir="ltr">
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Tue, Apr 23, 2019 at 7:46
AM Samuel Pitoiset <<a
href="mailto:samuel.pitoiset@gmail.com"
moz-do-not-send="true">samuel.pitoiset@gmail.com</a>>
wrote:<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left:1px solid
rgb(204,204,204);padding-left:1ex"><br>
On 4/23/19 10:45 AM, Bas Nieuwenhuizen wrote:<br>
> On Tue, Apr 23, 2019 at 9:35 AM Samuel Pitoiset<br>
> <<a href="mailto:samuel.pitoiset@gmail.com"
target="_blank" moz-do-not-send="true">samuel.pitoiset@gmail.com</a>>
wrote:<br>
>> Signed-off-by: Samuel Pitoiset <<a
href="mailto:samuel.pitoiset@gmail.com" target="_blank"
moz-do-not-send="true">samuel.pitoiset@gmail.com</a>><br>
>> ---<br>
>> src/amd/vulkan/radv_shader.c | 2
+-<br>
>> src/compiler/nir/nir.h | 3
++-<br>
>> src/compiler/nir/nir_opt_if.c | 17
++++++++++-------<br>
>> src/freedreno/ir3/ir3_nir.c | 2
+-<br>
>> src/gallium/auxiliary/nir/tgsi_to_nir.c | 2
+-<br>
>> src/gallium/drivers/freedreno/a2xx/ir2_nir.c | 2
+-<br>
>> src/gallium/drivers/radeonsi/si_shader_nir.c | 2
+-<br>
>> src/intel/compiler/brw_nir.c | 2
+-<br>
>> src/mesa/state_tracker/st_glsl_to_nir.cpp | 2
+-<br>
>> 9 files changed, 19 insertions(+), 15
deletions(-)<br>
>><br>
>> diff --git a/src/amd/vulkan/radv_shader.c
b/src/amd/vulkan/radv_shader.c<br>
>> index 13f1f9aa9dc..54a4e732230 100644<br>
>> --- a/src/amd/vulkan/radv_shader.c<br>
>> +++ b/src/amd/vulkan/radv_shader.c<br>
>> @@ -158,7 +158,7 @@ radv_optimize_nir(struct
nir_shader *shader, bool optimize_conservatively,<br>
>> NIR_PASS(progress, shader,
nir_opt_remove_phis);<br>
>> NIR_PASS(progress,
shader, nir_opt_dce);<br>
>> }<br>
>> - NIR_PASS(progress, shader,
nir_opt_if, true);<br>
>> + NIR_PASS(progress, shader,
nir_opt_if, true, false);<br>
>> NIR_PASS(progress, shader,
nir_opt_dead_cf);<br>
>> NIR_PASS(progress, shader,
nir_opt_cse);<br>
>> NIR_PASS(progress, shader,
nir_opt_peephole_select, 8, true, true);<br>
>> diff --git a/src/compiler/nir/nir.h
b/src/compiler/nir/nir.h<br>
>> index 7d2062d3691..d7506d6ddd1 100644<br>
>> --- a/src/compiler/nir/nir.h<br>
>> +++ b/src/compiler/nir/nir.h<br>
>> @@ -3474,7 +3474,8 @@ bool nir_opt_gcm(nir_shader
*shader, bool value_number);<br>
>><br>
>> bool nir_opt_idiv_const(nir_shader *shader,
unsigned min_bit_size);<br>
>><br>
>> -bool nir_opt_if(nir_shader *shader, bool
aggressive_last_continue);<br>
>> +bool nir_opt_if(nir_shader *shader, bool
aggressive_last_continue,<br>
>> + bool skip_alu_of_phi);<br>
> Can we have a flag for this instead (e.g. something
like<br>
> nir_opt_if_skip_alu_of_phi)? I think have a function
with a bunch of<br>
> bools is less than ideal as you can't see at the
calling site what is<br>
> for what arg.<br>
Yes, that seems better to me.<br>
</blockquote>
<div><br>
</div>
<div>This is the worst kind of hack all around. We're making
NIR more complicated and adding a flag to disable a useful
and correct piece of an optimization, not because it causes
a perf regression but because the back-end compiler is
broken and this is easier than fixing it properly.
Seriously? Can't we just fix the LLVM back-end? Or, if
this optimization is actually doing something wrong, fix
it? Or maybe actually figure out what pattern is causing
LLVM to fall over and have a hack in your NIR -> LLVM
pass? On the list of "good ways to fix this problem", this
seems to be pretty far down if it hasn't fallen off the
bottom.<br>
</div>
</div>
</div>
</blockquote>
<p>Best hack of the month? :-)</p>
<p>As discussed over IRC, this is definitely not the best solution,
I don't like it either as I said.</p>
<p>I will work on a different solution maybe in our NIR->LLVM
pass.<br>
</p>
<p>The aggressive_last_continue option should also be removed (make
it default?).<br>
</p>
<blockquote type="cite"
cite="mid:CAOFGe94RBM09CMhhYpe0cgTpdQdRDN3y2VmGwFExpXpWhuvW8w@mail.gmail.com">
<div dir="ltr">
<div class="gmail_quote">
<div><br>
</div>
<div>--Jason<br>
</div>
<div> </div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left:1px solid
rgb(204,204,204);padding-left:1ex">
>> bool nir_opt_intrinsics(nir_shader *shader);<br>
>><br>
>> diff --git a/src/compiler/nir/nir_opt_if.c
b/src/compiler/nir/nir_opt_if.c<br>
>> index f674185f1e2..149b3bd1659 100644<br>
>> --- a/src/compiler/nir/nir_opt_if.c<br>
>> +++ b/src/compiler/nir/nir_opt_if.c<br>
>> @@ -1385,7 +1385,8 @@ opt_if_cf_list(nir_builder
*b, struct exec_list *cf_list,<br>
>> * not do anything to cause the metadata to
become invalid.<br>
>> */<br>
>> static bool<br>
>> -opt_if_safe_cf_list(nir_builder *b, struct
exec_list *cf_list)<br>
>> +opt_if_safe_cf_list(nir_builder *b, struct
exec_list *cf_list,<br>
>> + bool skip_alu_of_phi)<br>
>> {<br>
>> bool progress = false;<br>
>> foreach_list_typed(nir_cf_node, cf_node, node,
cf_list) {<br>
>> @@ -1395,16 +1396,17 @@
opt_if_safe_cf_list(nir_builder *b, struct exec_list
*cf_list)<br>
>><br>
>> case nir_cf_node_if: {<br>
>> nir_if *nif =
nir_cf_node_as_if(cf_node);<br>
>> - progress |= opt_if_safe_cf_list(b,
&nif->then_list);<br>
>> - progress |= opt_if_safe_cf_list(b,
&nif->else_list);<br>
>> + progress |= opt_if_safe_cf_list(b,
&nif->then_list, skip_alu_of_phi);<br>
>> + progress |= opt_if_safe_cf_list(b,
&nif->else_list, skip_alu_of_phi);<br>
>> progress |=
opt_if_evaluate_condition_use(b, nif);<br>
>> break;<br>
>> }<br>
>><br>
>> case nir_cf_node_loop: {<br>
>> nir_loop *loop =
nir_cf_node_as_loop(cf_node);<br>
>> - progress |= opt_if_safe_cf_list(b,
&loop->body);<br>
>> - progress |= opt_split_alu_of_phi(b,
loop);<br>
>> + progress |= opt_if_safe_cf_list(b,
&loop->body, skip_alu_of_phi);<br>
>> + if (!skip_alu_of_phi)<br>
>> + progress |= opt_split_alu_of_phi(b,
loop);<br>
>> break;<br>
>> }<br>
>><br>
>> @@ -1417,7 +1419,8 @@
opt_if_safe_cf_list(nir_builder *b, struct exec_list
*cf_list)<br>
>> }<br>
>><br>
>> bool<br>
>> -nir_opt_if(nir_shader *shader, bool
aggressive_last_continue)<br>
>> +nir_opt_if(nir_shader *shader, bool
aggressive_last_continue,<br>
>> + bool skip_alu_of_phi)<br>
>> {<br>
>> bool progress = false;<br>
>><br>
>> @@ -1430,7 +1433,7 @@ nir_opt_if(nir_shader
*shader, bool aggressive_last_continue)<br>
>><br>
>> nir_metadata_require(function->impl,
nir_metadata_block_index |<br>
>>
nir_metadata_dominance);<br>
>> - progress = opt_if_safe_cf_list(&b,
&function->impl->body);<br>
>> + progress = opt_if_safe_cf_list(&b,
&function->impl->body, skip_alu_of_phi);<br>
>> nir_metadata_preserve(function->impl,
nir_metadata_block_index |<br>
>>
nir_metadata_dominance);<br>
>><br>
>> diff --git a/src/freedreno/ir3/ir3_nir.c
b/src/freedreno/ir3/ir3_nir.c<br>
>> index 76230e3be50..1bec3c030a9 100644<br>
>> --- a/src/freedreno/ir3/ir3_nir.c<br>
>> +++ b/src/freedreno/ir3/ir3_nir.c<br>
>> @@ -147,7 +147,7 @@ ir3_optimize_loop(nir_shader
*s)<br>
>> OPT(s, nir_copy_prop);<br>
>> OPT(s, nir_opt_dce);<br>
>> }<br>
>> - progress |= OPT(s, nir_opt_if,
false);<br>
>> + progress |= OPT(s, nir_opt_if,
false, false);<br>
>> progress |= OPT(s,
nir_opt_remove_phis);<br>
>> progress |= OPT(s, nir_opt_undef);<br>
>><br>
>> diff --git
a/src/gallium/auxiliary/nir/tgsi_to_nir.c
b/src/gallium/auxiliary/nir/tgsi_to_nir.c<br>
>> index c55e8b84a41..6b40bff1f73 100644<br>
>> --- a/src/gallium/auxiliary/nir/tgsi_to_nir.c<br>
>> +++ b/src/gallium/auxiliary/nir/tgsi_to_nir.c<br>
>> @@ -2066,7 +2066,7 @@ ttn_optimize_nir(nir_shader
*nir, bool scalar)<br>
>> NIR_PASS(progress, nir, nir_opt_dce);<br>
>> }<br>
>><br>
>> - NIR_PASS(progress, nir, nir_opt_if, false);<br>
>> + NIR_PASS(progress, nir, nir_opt_if, false,
false);<br>
>> NIR_PASS(progress, nir, nir_opt_dead_cf);<br>
>> NIR_PASS(progress, nir, nir_opt_cse);<br>
>> NIR_PASS(progress, nir,
nir_opt_peephole_select, 8, true, true);<br>
>> diff --git
a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c
b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c<br>
>> index ee348ca6a93..3522971e435 100644<br>
>> --- a/src/gallium/drivers/freedreno/a2xx/ir2_nir.c<br>
>> +++ b/src/gallium/drivers/freedreno/a2xx/ir2_nir.c<br>
>> @@ -94,7 +94,7 @@ ir2_optimize_loop(nir_shader *s)<br>
>> OPT(s, nir_opt_dce);<br>
>> }<br>
>> progress |= OPT(s,
nir_opt_loop_unroll, nir_var_all);<br>
>> - progress |= OPT(s, nir_opt_if,
false);<br>
>> + progress |= OPT(s, nir_opt_if,
false, false);<br>
>> progress |= OPT(s,
nir_opt_remove_phis);<br>
>> progress |= OPT(s, nir_opt_undef);<br>
>><br>
>> diff --git
a/src/gallium/drivers/radeonsi/si_shader_nir.c
b/src/gallium/drivers/radeonsi/si_shader_nir.c<br>
>> index 5a925f19e09..7f1fe4ba2e9 100644<br>
>> --- a/src/gallium/drivers/radeonsi/si_shader_nir.c<br>
>> +++ b/src/gallium/drivers/radeonsi/si_shader_nir.c<br>
>> @@ -879,7 +879,7 @@ si_lower_nir(struct
si_shader_selector* sel)<br>
>> NIR_PASS(progress,
sel->nir, nir_copy_prop);<br>
>> NIR_PASS(progress,
sel->nir, nir_opt_dce);<br>
>> }<br>
>> - NIR_PASS(progress, sel->nir,
nir_opt_if, true);<br>
>> + NIR_PASS(progress, sel->nir,
nir_opt_if, true, false);<br>
>> NIR_PASS(progress, sel->nir,
nir_opt_dead_cf);<br>
>> NIR_PASS(progress, sel->nir,
nir_opt_cse);<br>
>> NIR_PASS(progress, sel->nir,
nir_opt_peephole_select, 8, true, true);<br>
>> diff --git a/src/intel/compiler/brw_nir.c
b/src/intel/compiler/brw_nir.c<br>
>> index e0a393fc298..ba911049ce3 100644<br>
>> --- a/src/intel/compiler/brw_nir.c<br>
>> +++ b/src/intel/compiler/brw_nir.c<br>
>> @@ -607,7 +607,7 @@ brw_nir_optimize(nir_shader
*nir, const struct brw_compiler *compiler,<br>
>> OPT(nir_copy_prop);<br>
>> OPT(nir_opt_dce);<br>
>> }<br>
>> - OPT(nir_opt_if, false);<br>
>> + OPT(nir_opt_if, false, false);<br>
>> if
(nir->options->max_unroll_iterations != 0) {<br>
>> OPT(nir_opt_loop_unroll, indirect_mask);<br>
>> }<br>
>> diff --git
a/src/mesa/state_tracker/st_glsl_to_nir.cpp
b/src/mesa/state_tracker/st_glsl_to_nir.cpp<br>
>> index 97b2831b880..3f1f78e875b 100644<br>
>> --- a/src/mesa/state_tracker/st_glsl_to_nir.cpp<br>
>> +++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp<br>
>> @@ -324,7 +324,7 @@ st_nir_opts(nir_shader *nir,
bool scalar)<br>
>> NIR_PASS(progress, nir, nir_copy_prop);<br>
>> NIR_PASS(progress, nir, nir_opt_dce);<br>
>> }<br>
>> - NIR_PASS(progress, nir, nir_opt_if, false);<br>
>> + NIR_PASS(progress, nir, nir_opt_if, false,
false);<br>
>> NIR_PASS(progress, nir, nir_opt_dead_cf);<br>
>> NIR_PASS(progress, nir, nir_opt_cse);<br>
>> NIR_PASS(progress, nir,
nir_opt_peephole_select, 8, true, true);<br>
>> --<br>
>> 2.21.0<br>
>><br>
>> _______________________________________________<br>
>> mesa-dev mailing list<br>
>> <a href="mailto:mesa-dev@lists.freedesktop.org"
target="_blank" moz-do-not-send="true">mesa-dev@lists.freedesktop.org</a><br>
>> <a
href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev"
rel="noreferrer" target="_blank" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a><br>
_______________________________________________<br>
mesa-dev mailing list<br>
<a href="mailto:mesa-dev@lists.freedesktop.org"
target="_blank" moz-do-not-send="true">mesa-dev@lists.freedesktop.org</a><br>
<a
href="https://lists.freedesktop.org/mailman/listinfo/mesa-dev"
rel="noreferrer" target="_blank" moz-do-not-send="true">https://lists.freedesktop.org/mailman/listinfo/mesa-dev</a></blockquote>
</div>
</div>
</blockquote>
</body>
</html>