[Mesa-dev] [PATCH 3/4] nir: support to flatten_all in peephole-select

[Jason Ekstrand]

On Tue, Mar 31, 2015 at 3:57 PM, Rob Clark <robdclark at gmail.com> wrote:
> From: Rob Clark <robclark at freedesktop.org>
>
> Freedreno and vc4 want this behavior for the time being (until we have
> real flow control).  Even after that, we probably want to turn this into
> some sort of driver tunable threshold, since for at least some hardware,
> reasonably large if/else is best flattend rather than having divergent
> flow control.
>
> NOTE: wasn't sure about some of the additional restrictions in
> block_check_for_allowed_instrs()..  are there some other cases where
> I might need to fix things up in order to be guaranteed to be able to
> flatten?
>
> NOTE: drop vc4 hunk if this is merged first, ofc
>
> Signed-off-by: Rob Clark <robclark at freedesktop.org>
> ---
>  src/gallium/drivers/vc4/vc4_program.c    |  2 +-
>  src/glsl/nir/nir.h                       |  2 +-
>  src/glsl/nir/nir_opt_peephole_select.c   | 21 +++++++++++++++------
>  src/mesa/drivers/dri/i965/brw_fs_nir.cpp |  2 +-
>  4 files changed, 18 insertions(+), 9 deletions(-)
>
> diff --git a/src/gallium/drivers/vc4/vc4_program.c b/src/gallium/drivers/vc4/vc4_program.c
> index db51665..09896ce 100644
> --- a/src/gallium/drivers/vc4/vc4_program.c
> +++ b/src/gallium/drivers/vc4/vc4_program.c
> @@ -1673,7 +1673,7 @@ vc4_optimize_nir(struct nir_shader *s)
>                  progress = nir_copy_prop(s) || progress;
>                  progress = nir_opt_dce(s) || progress;
>                  progress = nir_opt_cse(s) || progress;
> -                progress = nir_opt_peephole_select(s) || progress;
> +                progress = nir_opt_peephole_select(s, true) || progress;
>                  progress = nir_opt_algebraic(s) || progress;
>                  progress = nir_opt_constant_folding(s) || progress;
>          } while (progress);
> diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
> index 74927e5..cd03d6b 100644
> --- a/src/glsl/nir/nir.h
> +++ b/src/glsl/nir/nir.h
> @@ -1631,7 +1631,7 @@ bool nir_opt_dce(nir_shader *shader);
>
>  void nir_opt_gcm(nir_shader *shader);
>
> -bool nir_opt_peephole_select(nir_shader *shader);
> +bool nir_opt_peephole_select(nir_shader *shader, bool flatten_all);
>  bool nir_opt_peephole_ffma(nir_shader *shader);
>
>  bool nir_opt_remove_phis(nir_shader *shader);
> diff --git a/src/glsl/nir/nir_opt_peephole_select.c b/src/glsl/nir/nir_opt_peephole_select.c
> index b89451b..e48fc7a 100644
> --- a/src/glsl/nir/nir_opt_peephole_select.c
> +++ b/src/glsl/nir/nir_opt_peephole_select.c
> @@ -44,10 +44,16 @@
>   * whose only use is one of the following phi nodes.  This happens all the
>   * time when the SSA form comes from a conditional assignment with a
>   * swizzle.
> + *
> + * If 'flatten_all' is true, then flatten *all* if/else into one block
> + * and use select to pick the winners.  This is useful for drivers that
> + * do not (yet) have proper flow control.  Eventually we probably want
> + * to make this a more clever driver tunable threshold.
>   */
>
>  struct peephole_select_state {
>     void *mem_ctx;
> +   bool flatten_all;
>     bool progress;
>  };
>
> @@ -150,9 +156,10 @@ nir_opt_peephole_select_block(nir_block *block, void *void_state)
>     nir_block *else_block = nir_cf_node_as_block(else_node);
>
>     /* ... and those blocks must only contain "allowed" instructions. */
> -   if (!block_check_for_allowed_instrs(then_block) ||
> -       !block_check_for_allowed_instrs(else_block))
> -      return true;
> +   if (!state->flatten_all)
> +      if (!block_check_for_allowed_instrs(then_block) ||
> +          !block_check_for_allowed_instrs(else_block))
> +         return true;

This isn't valid.  We can't blindly flatten if either list of blocks
contains an instruction that has side-effects.  In order to do that,
we would have to add predication back into NIR.

>     /* At this point, we know that the previous CFG node is an if-then
>      * statement containing only moves to phi nodes in this block.  We can
> @@ -217,11 +224,12 @@ nir_opt_peephole_select_block(nir_block *block, void *void_state)
>  }
>
>  static bool
> -nir_opt_peephole_select_impl(nir_function_impl *impl)
> +nir_opt_peephole_select_impl(nir_function_impl *impl, bool flatten_all)
>  {
>     struct peephole_select_state state;
>
>     state.mem_ctx = ralloc_parent(impl);
> +   state.flatten_all = flatten_all;
>     state.progress = false;
>
>     nir_foreach_block(impl, nir_opt_peephole_select_block, &state);
> @@ -233,13 +241,14 @@ nir_opt_peephole_select_impl(nir_function_impl *impl)
>  }
>
>  bool
> -nir_opt_peephole_select(nir_shader *shader)
> +nir_opt_peephole_select(nir_shader *shader, bool flatten_all)

As matt said, boolean arguments aren't great.  I also don't like
arbitrary bitfields.  We do, however, have a shader params struct.
You could add a "flatten" flag to it I guess.  However, doing so kind
of implies that we actually flatten everything which (as stated above)
we can't here.
--Jason

>  {
>     bool progress = false;
>
>     nir_foreach_overload(shader, overload) {
>        if (overload->impl)
> -         progress |= nir_opt_peephole_select_impl(overload->impl);
> +         progress |= nir_opt_peephole_select_impl(overload->impl,
> +                                                  flatten_all);
>     }
>
>     return progress;
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> index 0b8ed1a..7bd7cd3 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
> @@ -47,7 +47,7 @@ nir_optimize(nir_shader *nir)
>        nir_validate_shader(nir);
>        progress |= nir_opt_cse(nir);
>        nir_validate_shader(nir);
> -      progress |= nir_opt_peephole_select(nir);
> +      progress |= nir_opt_peephole_select(nir, false);
>        nir_validate_shader(nir);
>        progress |= nir_opt_algebraic(nir);
>        nir_validate_shader(nir);
> --
> 2.1.0
>