[Mesa-dev] [PATCH 07/16] i965/fs: Support coalescing registers of size > 1.

[Jordan Justen]

On Thu, Dec 19, 2013 at 1:40 PM, Matt Turner <mattst88 at gmail.com> wrote:
> total instructions in shared programs: 1550048 -> 1549880 (-0.01%)
> instructions in affected programs:     1896 -> 1728 (-8.86%)
> ---
>  src/mesa/drivers/dri/i965/brw_fs.cpp | 80 ++++++++++++++++++++++++++----------
>  1 file changed, 58 insertions(+), 22 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
> index 1a16f4e..39041e3 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
> @@ -2265,6 +2265,12 @@ fs_visitor::register_coalesce()
>
>     calculate_live_intervals();
>
> +   int src_size = 0;
> +   int channels_remaining = 0;
> +   int reg_from = -1, reg_to = -1;
> +   int reg_to_offset[MAX_SAMPLER_MESSAGE_SIZE];
> +   fs_inst *mov[MAX_SAMPLER_MESSAGE_SIZE];
> +
>     foreach_list_safe(node, &this->instructions) {
>        fs_inst *inst = (fs_inst *)node;
>
> @@ -2276,11 +2282,14 @@ fs_visitor::register_coalesce()
>           inst->src[0].abs ||
>           inst->src[0].smear != -1 ||
>           inst->dst.file != GRF ||
> -         inst->dst.type != inst->src[0].type ||
> -         virtual_grf_sizes[inst->src[0].reg] != 1) {
> +         inst->dst.type != inst->src[0].type) {
>          continue;
>        }
>
> +      if (virtual_grf_sizes[inst->src[0].reg] >
> +          virtual_grf_sizes[inst->dst.reg])
> +         continue;

Why isn't this != rather than >?

>        int var_from = live_intervals->var_from_reg(&inst->src[0]);
>        int var_to = live_intervals->var_from_reg(&inst->dst);
>
> @@ -2288,31 +2297,58 @@ fs_visitor::register_coalesce()
>            !inst->dst.equals(inst->src[0]))
>           continue;
>
> -      int reg_from = inst->src[0].reg;
> -      assert(inst->src[0].reg_offset == 0);
> -      int reg_to = inst->dst.reg;
> -      int reg_to_offset = inst->dst.reg_offset;
> +      if (reg_from != inst->src[0].reg) {
> +         reg_from = inst->src[0].reg;
> +
> +         src_size = virtual_grf_sizes[inst->src[0].reg];
> +         assert(src_size <= MAX_SAMPLER_MESSAGE_SIZE);
> +
> +         channels_remaining = src_size;
> +         memset(mov, 0, sizeof(mov));
> +
> +         reg_to = inst->dst.reg;
> +      }
> +
> +      if (reg_to != inst->dst.reg)
> +         continue;
> +
> +      const int offset = inst->src[0].reg_offset;
> +      reg_to_offset[offset] = inst->dst.reg_offset;
> +      mov[offset] = inst;
> +      channels_remaining--;
> +
> +      if (channels_remaining)
> +         continue;
> +
> +      for (int i = 0; i < src_size; i++) {
> +         if (mov[i])
> +            mov[i]->remove();
> +      }
>
>        foreach_list(node, &this->instructions) {
>          fs_inst *scan_inst = (fs_inst *)node;
>
> -        if (scan_inst->dst.file == GRF &&
> -            scan_inst->dst.reg == reg_from) {
> -           scan_inst->dst.reg = reg_to;
> -           scan_inst->dst.reg_offset = reg_to_offset;
> -        }
> -        for (int i = 0; i < 3; i++) {
> -           if (scan_inst->src[i].file == GRF &&
> -               scan_inst->src[i].reg == reg_from) {
> -              scan_inst->src[i].reg = reg_to;
> -              scan_inst->src[i].reg_offset = reg_to_offset;
> -           }
> -        }
> -      }
> +         for (int i = 0; i < src_size; i++) {
> +            if (mov[i]) {

Shouldn't this always be taken?

> +               if (scan_inst->dst.file == GRF &&
> +                   scan_inst->dst.reg == reg_from &&
> +                   scan_inst->dst.reg_offset == i) {
> +                  scan_inst->dst.reg = reg_to;
> +                  scan_inst->dst.reg_offset = reg_to_offset[i];
> +               }
> +               for (int j = 0; j < 3; j++) {
> +                  if (scan_inst->src[j].file == GRF &&
> +                      scan_inst->src[j].reg == reg_from &&
> +                      scan_inst->src[j].reg_offset == i) {
> +                     scan_inst->src[j].reg = reg_to;
> +                     scan_inst->src[j].reg_offset = reg_to_offset[i];
> +                  }
> +               }
>
> -      inst->remove();
> -      progress = true;
> -      continue;
> +               progress = true;
> +            }
> +         }
> +      }

Couldn't we assign progress = true here instead? Ie, outside 2 of the
3 nested for loops.

>     }
>
>     if (progress)
> --
> 1.8.3.2
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev