[Mesa-dev] [PATCH 3/9] nir/opt_peephole_select: Don't peephole_select expensive math instructions

Jason Ekstrand jason at jlekstrand.net
Tue Oct 9 18:15:28 UTC 2018


Drive-by comment: At some point, maybe we just want to replace all these 
booleans with a function pointer that takes an instruction and returns a 
bool.  That way the driver can define "expensive".  Do with that what you will.

On October 9, 2018 11:50:45 Ian Romanick <idr at freedesktop.org> wrote:

> On 10/08/2018 01:34 PM, Thomas Helland wrote:
>> Den tor. 30. aug. 2018 kl. 07:37 skrev Ian Romanick <idr at freedesktop.org>:
>>>
>>> From: Ian Romanick <ian.d.romanick at intel.com>
>>>
>>> On some GPUs, especially older Intel GPUs, some math instructions are
>>> very expensive.  On those architectures, don't reduce flow control to a
>>> csel if one of the branches contains one of these expensive math
>>> instructions.
>>>
>>> This prevents a bunch of cycle count regressions on pre-Gen6 platforms
>>> with a later patch (intel/compiler: More peephole select for pre-Gen6).
>>>
>>> Signed-off-by: Ian Romanick <ian.d.romanick at intel.com>
>>> ---
>>>  src/amd/vulkan/radv_shader.c                 |  2 +-
>>>  src/broadcom/compiler/nir_to_vir.c           |  2 +-
>>>  src/compiler/nir/nir.h                       |  2 +-
>>>  src/compiler/nir/nir_opt_peephole_select.c   | 46 +++++++++++++++++++++++-----
>>>  src/gallium/drivers/freedreno/ir3/ir3_nir.c  |  2 +-
>>>  src/gallium/drivers/radeonsi/si_shader_nir.c |  2 +-
>>>  src/gallium/drivers/vc4/vc4_program.c        |  2 +-
>>>  src/intel/compiler/brw_nir.c                 |  4 +--
>>>  src/mesa/state_tracker/st_glsl_to_nir.cpp    |  2 +-
>>>  9 files changed, 47 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
>>> index 632512db09b..c8d502a9e3a 100644
>>> --- a/src/amd/vulkan/radv_shader.c
>>> +++ b/src/amd/vulkan/radv_shader.c
>>> @@ -143,7 +143,7 @@ radv_optimize_nir(struct nir_shader *shader, bool 
>>> optimize_conservatively)
>>>                  NIR_PASS(progress, shader, nir_opt_if);
>>>                  NIR_PASS(progress, shader, nir_opt_dead_cf);
>>>                  NIR_PASS(progress, shader, nir_opt_cse);
>>> -                NIR_PASS(progress, shader, nir_opt_peephole_select, 8, true);
>>> +                NIR_PASS(progress, shader, nir_opt_peephole_select, 8, 
>>> true, true);
>>>                  NIR_PASS(progress, shader, nir_opt_algebraic);
>>>                  NIR_PASS(progress, shader, nir_opt_constant_folding);
>>>                  NIR_PASS(progress, shader, nir_opt_undef);
>>> diff --git a/src/broadcom/compiler/nir_to_vir.c 
>>> b/src/broadcom/compiler/nir_to_vir.c
>>> index 0d23cea4d5b..ec0ff4b907a 100644
>>> --- a/src/broadcom/compiler/nir_to_vir.c
>>> +++ b/src/broadcom/compiler/nir_to_vir.c
>>> @@ -1210,7 +1210,7 @@ v3d_optimize_nir(struct nir_shader *s)
>>>                  NIR_PASS(progress, s, nir_opt_dce);
>>>                  NIR_PASS(progress, s, nir_opt_dead_cf);
>>>                  NIR_PASS(progress, s, nir_opt_cse);
>>> -                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true);
>>> +                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
>>>                  NIR_PASS(progress, s, nir_opt_algebraic);
>>>                  NIR_PASS(progress, s, nir_opt_constant_folding);
>>>                  NIR_PASS(progress, s, nir_opt_undef);
>>> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
>>> index 67fa46d5557..feb69be6b59 100644
>>> --- a/src/compiler/nir/nir.h
>>> +++ b/src/compiler/nir/nir.h
>>> @@ -3003,7 +3003,7 @@ bool nir_opt_move_comparisons(nir_shader *shader);
>>>  bool nir_opt_move_load_ubo(nir_shader *shader);
>>>
>>>  bool nir_opt_peephole_select(nir_shader *shader, unsigned limit,
>>> -                             bool indirect_load_ok);
>>> +                             bool indirect_load_ok, bool expensive_alu_ok);
>>>
>>>  bool nir_opt_remove_phis_impl(nir_function_impl *impl);
>>>  bool nir_opt_remove_phis(nir_shader *shader);
>>> diff --git a/src/compiler/nir/nir_opt_peephole_select.c 
>>> b/src/compiler/nir/nir_opt_peephole_select.c
>>> index 6808d3eda6c..09b55f3739e 100644
>>> --- a/src/compiler/nir/nir_opt_peephole_select.c
>>> +++ b/src/compiler/nir/nir_opt_peephole_select.c
>>> @@ -59,7 +59,8 @@
>>>
>>>  static bool
>>>  block_check_for_allowed_instrs(nir_block *block, unsigned *count,
>>> -                               bool alu_ok, bool indirect_load_ok)
>>> +                               bool alu_ok, bool indirect_load_ok,
>>> +                               bool expensive_alu_ok)
>>>  {
>>>     nir_foreach_instr(instr, block) {
>>>        switch (instr->type) {
>>> @@ -117,6 +118,25 @@ block_check_for_allowed_instrs(nir_block *block, 
>>> unsigned *count,
>>>           case nir_op_vec3:
>>>           case nir_op_vec4:
>>>              break;
>>> +
>>> +         case nir_op_fcos:
>>> +         case nir_op_fdiv:
>>> +         case nir_op_fexp2:
>>> +         case nir_op_flog2:
>>> +         case nir_op_fmod:
>>> +         case nir_op_fpow:
>>> +         case nir_op_frcp:
>>> +         case nir_op_frem:
>>> +         case nir_op_frsq:
>>> +         case nir_op_fsin:
>>> +         case nir_op_idiv:
>>> +         case nir_op_irem:
>>> +         case nir_op_udiv:
>>> +            if (!alu_ok || !expensive_alu_ok)
>>> +               return false;
>>> +
>>> +            break;
>>> +
>>>           default:
>>>              if (!alu_ok) {
>>>                 /* It must be a move-like operation. */
>>> @@ -160,7 +180,8 @@ block_check_for_allowed_instrs(nir_block *block, 
>>> unsigned *count,
>>>
>>>  static bool
>>>  nir_opt_peephole_select_block(nir_block *block, nir_shader *shader,
>>> -                              unsigned limit, bool indirect_load_ok)
>>> +                              unsigned limit, bool indirect_load_ok,
>>> +                              bool expensive_alu_ok)
>>>  {
>>>     if (nir_cf_node_is_first(&block->cf_node))
>>>        return false;
>>> @@ -180,10 +201,17 @@ nir_opt_peephole_select_block(nir_block *block, 
>>> nir_shader *shader,
>>>
>>>     /* ... and those blocks must only contain "allowed" instructions. */
>>>     unsigned count = 0;
>>> +#if 1
>>>     if (!block_check_for_allowed_instrs(then_block, &count, limit != 0,
>>> -                                       indirect_load_ok) ||
>>> +                                       indirect_load_ok, expensive_alu_ok) ||
>>>         !block_check_for_allowed_instrs(else_block, &count, limit != 0,
>>> -                                       indirect_load_ok))
>>> +                                       indirect_load_ok, expensive_alu_ok))
>>> +#else
>>> +   if (!block_check_for_allowed_instrs(then_block, &count, shader->info.stage,
>>> +                                       limit != 0, indirect_load_ok, 
>>> expensive_alu_ok) ||
>>> +       !block_check_for_allowed_instrs(else_block, &count, shader->info.stage,
>>> +                                       limit != 0, indirect_load_ok, 
>>> expensive_alu_ok))
>>> +#endif
>>
>> Leftover testing stuff?
>
> Yes. :)  I was experimenting with doing things differently for the
> scalar and vector units on older Intel GPUs.
>
>> I like the idea of hiding expensive instructions in a branch.
>> However, I'm wondering if it might be an idea to let drivers
>> provide a callback with what instructions they want to allow?
>> If no other driver plan on doing something like this then I
>> guess this is OK. I would expect the "expensive" instructions
>> to be mostly the same subset on most architectures too.
>> Might want to get some input from others, but with the debug
>> stuff above removed this is.
>
> There are a few things that could be done with this pass or a related
> pass.  Before going to far down the expensive instruction path, I'd want
> to experiment with a pass that could rearrange things like:
>
> 	if (condition)
> 		x = sin(a);
> 	else
> 		x = sin(b);
>
> I'd also like experiment with taking into consideration whether or not
> 'condition' is uniform or not.  Some architectures may want to make
> different flattening choices for uniform vs. non-uniform.
>
> Unfortunately, shader-db is not a good tool for measuring the effects of
> this.
>
>> Reviewed-by: Thomas Helland<thomashelland90 at gmail.com>
>>
>>>        return false;
>>>
>>>     if (count > limit)
>>> @@ -250,14 +278,15 @@ nir_opt_peephole_select_block(nir_block *block, 
>>> nir_shader *shader,
>>>
>>>  static bool
>>>  nir_opt_peephole_select_impl(nir_function_impl *impl, unsigned limit,
>>> -                             bool indirect_load_ok)
>>> +                             bool indirect_load_ok, bool expensive_alu_ok)
>>>  {
>>>     nir_shader *shader = impl->function->shader;
>>>     bool progress = false;
>>>
>>>     nir_foreach_block_safe(block, impl) {
>>>        progress |= nir_opt_peephole_select_block(block, shader, limit,
>>> -                                                indirect_load_ok);
>>> +                                                indirect_load_ok,
>>> +                                                expensive_alu_ok);
>>>     }
>>>
>>>     if (progress)
>>> @@ -268,14 +297,15 @@ nir_opt_peephole_select_impl(nir_function_impl *impl, 
>>> unsigned limit,
>>>
>>>  bool
>>>  nir_opt_peephole_select(nir_shader *shader, unsigned limit,
>>> -                        bool indirect_load_ok)
>>> +                        bool indirect_load_ok, bool expensive_alu_ok)
>>>  {
>>>     bool progress = false;
>>>
>>>     nir_foreach_function(function, shader) {
>>>        if (function->impl)
>>>           progress |= nir_opt_peephole_select_impl(function->impl, limit,
>>> -                                                  indirect_load_ok);
>>> +                                                  indirect_load_ok,
>>> +                                                  expensive_alu_ok);
>>>     }
>>>
>>>     return progress;
>>> diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir.c 
>>> b/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>> index 5f66ef5d170..bb3bb73644a 100644
>>> --- a/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>> +++ b/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>> @@ -107,7 +107,7 @@ ir3_optimize_loop(nir_shader *s)
>>>                         progress |= OPT(s, nir_opt_gcm, true);
>>>                 else if (gcm == 2)
>>>                         progress |= OPT(s, nir_opt_gcm, false);
>>> -               progress |= OPT(s, nir_opt_peephole_select, 16, true);
>>> +               progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
>>>                 progress |= OPT(s, nir_opt_intrinsics);
>>>                 progress |= OPT(s, nir_opt_algebraic);
>>>                 progress |= OPT(s, nir_opt_constant_folding);
>>> diff --git a/src/gallium/drivers/radeonsi/si_shader_nir.c 
>>> b/src/gallium/drivers/radeonsi/si_shader_nir.c
>>> index 9a7a8264283..cb1e208be8f 100644
>>> --- a/src/gallium/drivers/radeonsi/si_shader_nir.c
>>> +++ b/src/gallium/drivers/radeonsi/si_shader_nir.c
>>> @@ -813,7 +813,7 @@ si_lower_nir(struct si_shader_selector* sel)
>>>                 NIR_PASS(progress, sel->nir, nir_opt_if);
>>>                 NIR_PASS(progress, sel->nir, nir_opt_dead_cf);
>>>                 NIR_PASS(progress, sel->nir, nir_opt_cse);
>>> -               NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8, true);
>>> +               NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8, 
>>> true, true);
>>>
>>>                 /* Needed for algebraic lowering */
>>>                 NIR_PASS(progress, sel->nir, nir_opt_algebraic);
>>> diff --git a/src/gallium/drivers/vc4/vc4_program.c 
>>> b/src/gallium/drivers/vc4/vc4_program.c
>>> index 39f7db9148c..fa2a32922f4 100644
>>> --- a/src/gallium/drivers/vc4/vc4_program.c
>>> +++ b/src/gallium/drivers/vc4/vc4_program.c
>>> @@ -1580,7 +1580,7 @@ vc4_optimize_nir(struct nir_shader *s)
>>>                  NIR_PASS(progress, s, nir_opt_dce);
>>>                  NIR_PASS(progress, s, nir_opt_dead_cf);
>>>                  NIR_PASS(progress, s, nir_opt_cse);
>>> -                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true);
>>> +                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
>>>                  NIR_PASS(progress, s, nir_opt_algebraic);
>>>                  NIR_PASS(progress, s, nir_opt_constant_folding);
>>>                  NIR_PASS(progress, s, nir_opt_undef);
>>> diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
>>> index 1d65107a93d..04d399fd0b4 100644
>>> --- a/src/intel/compiler/brw_nir.c
>>> +++ b/src/intel/compiler/brw_nir.c
>>> @@ -587,9 +587,9 @@ brw_nir_optimize(nir_shader *nir, const struct 
>>> brw_compiler *compiler,
>>>        const bool is_vec4_tessellation = !is_scalar &&
>>>           (nir->info.stage == MESA_SHADER_TESS_CTRL ||
>>>            nir->info.stage == MESA_SHADER_TESS_EVAL);
>>> -      OPT(nir_opt_peephole_select, 0, is_vec4_tessellation);
>>> +      OPT(nir_opt_peephole_select, 0, is_vec4_tessellation, false);
>>>        if (compiler->devinfo->gen >= 6)
>>> -         OPT(nir_opt_peephole_select, 1, is_vec4_tessellation);
>>> +         OPT(nir_opt_peephole_select, 1, is_vec4_tessellation, true);
>>>
>>>        OPT(nir_opt_intrinsics);
>>>        OPT(nir_opt_algebraic);
>>> diff --git a/src/mesa/state_tracker/st_glsl_to_nir.cpp 
>>> b/src/mesa/state_tracker/st_glsl_to_nir.cpp
>>> index bfcef3a293f..098a1e98833 100644
>>> --- a/src/mesa/state_tracker/st_glsl_to_nir.cpp
>>> +++ b/src/mesa/state_tracker/st_glsl_to_nir.cpp
>>> @@ -344,7 +344,7 @@ st_nir_opts(nir_shader *nir, bool scalar)
>>>        NIR_PASS(progress, nir, nir_opt_if);
>>>        NIR_PASS(progress, nir, nir_opt_dead_cf);
>>>        NIR_PASS(progress, nir, nir_opt_cse);
>>> -      NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true);
>>> +      NIR_PASS(progress, nir, nir_opt_peephole_select, 8, true, true);
>>>
>>>        NIR_PASS(progress, nir, nir_opt_algebraic);
>>>        NIR_PASS(progress, nir, nir_opt_constant_folding);
>>> --
>>> 2.14.4
>>>
>>> _______________________________________________
>>> mesa-dev mailing list
>>> mesa-dev at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev





More information about the mesa-dev mailing list