[Mesa-dev] [PATCH 13/25] i965/fs: Skip SIMD lowering destination zipping if possible.

[Jason Ekstrand]

On Fri, May 27, 2016 at 7:05 PM, Francisco Jerez <currojerez at riseup.net>
wrote:

> Skipping the temporary allocation and copy instructions is easy (just
> return dst), but the conditions used to find out whether the copy can
> be optimized out safely without breaking the program are rather
> complex: The destination must be exactly one component of at most the
> execution width of the lowered instruction, and all source regions of
> the instruction must be either fully disjoint from the destination or
> be aligned with it group by group.
> ---
>  src/mesa/drivers/dri/i965/brw_fs.cpp | 81
> ++++++++++++++++++++++++++++++++++++
>  1 file changed, 81 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp
> b/src/mesa/drivers/dri/i965/brw_fs.cpp
> index 8eff27e..5d26c68 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
> @@ -5054,6 +5054,78 @@ emit_unzip(const fs_builder &lbld, bblock_t *block,
> fs_inst *inst,
>  }
>
>  /**
> + * Return true if splitting out the group of channels of instruction \p
> inst
> + * given by lbld.group() requires allocating a temporary for the
> destination
> + * of the lowered instruction and copying the data back to the original
> + * destination region.
> + *
> + * Note that the result of this function may be different for different
> + * channel groups in cases where the overlap between a source and
> destination
> + * region is partial.
> + */
> +static inline bool
> +needs_dst_copy(const fs_builder &lbld, const fs_inst *inst)
> +{
> +   /* If the instruction writes more than one component we'll have to
> shuffle
> +    * the results of multiple lowered instructions in order to make sure
> that
> +    * they end up arranged correctly in the original destination region.
> +    */
> +   if (inst->regs_written * REG_SIZE >
> +       inst->dst.component_size(inst->exec_size))
> +      return true;
> +
> +   /* If the lowered execution size is larger than the original the
> result of
> +    * the instruction won't fit in the original destination, so we'll
> have to
> +    * allocate a temporary in any case.
> +    */
> +   if (lbld.dispatch_width() > inst->exec_size)
> +      return true;
> +
> +   for (unsigned i = 0; i < inst->sources; i++) {
> +      /* Index of this lowered instruction and total number of lowered
> +       * instructions.
> +       */
> +      const unsigned j = lbld.group() / lbld.dispatch_width();
> +      const unsigned n = DIV_ROUND_UP(inst->exec_size,
> lbld.dispatch_width());
> +
> +      /* Specified channels from the source and destination regions. */
> +      const fs_reg src = horiz_offset(inst->src[i], lbld.group());
> +      const fs_reg dst = horiz_offset(inst->dst, lbld.group());
> +
> +      /* Lowered component size of the source and destination regions. */
> +      const unsigned dsrc =
> inst->src[i].component_size(lbld.dispatch_width());
> +      const unsigned ddst =
> inst->dst.component_size(lbld.dispatch_width());
> +
> +      /* If we already made a copy of the source for other reasons there
> won't
> +       * be any overlap with the destination.
> +       */
> +      if (!is_periodic(inst->src[i], lbld.dispatch_width()) &&
> +          inst->components_read(i) != 1)
>

might be good to have a needs_src_copy function to keep the logic in one
place.


> +         continue;
> +
> +      /* We need a copy if the specified group of channels of the
> destination
> +       * overlaps either the previous or subsequent groups of the
> source.  We
> +       * don't care whether the destination of a given lowered instruction
> +       * overlaps its own source because there's arguably no
> +       * source/destination hazard for this opcode if the source and
> +       * destination of the unlowered instruction were already
> overlapping.
> +       *
> +       * Note that it might be possible to avoid emitting a copy in more
> cases
> +       * by making assumptions about the ordering in which the lowered
> +       * instructions will be emitted (ugh!), but this has the advantage
> that
> +       * the lowered instructions will be fully independent from each
> other
> +       * regardless of the form of overlap so they can be scheduled
> freely.
> +       */
> +      if (regions_overlap(dst, ddst, inst->src[i], dsrc * j) ||
> +          regions_overlap(dst, ddst, horiz_offset(src,
> lbld.dispatch_width()),
> +                           dsrc * (n - j - 1)))
> +         return true;
>

Wow, this is annoyingly complicated.  It also appears to be correct.  Would
it be sufficient to simply use the following?

if (regions_overlap(inst->dst, inst->regs_written * REG_SIZE, inst->src[i],
inst->regs_read() * REG_SIZE) &&
    !inst->dst.equals(inst->src[i]))
   return true;

It's much simpler and covers the two cases we really care about.  If that's
not sufficient, what you have above does look correct.

--Jason


> +   }
> +
> +   return false;
> +}
> +
> +/**
>   * Insert data from a packed temporary into the channel group given by
>   * lbld.group() of the destination region of instruction \p inst and
> return
>   * the temporary as result.  If any copy instructions are required they
> will
> @@ -5073,6 +5145,8 @@ emit_zip(const fs_builder &lbld, bblock_t *block,
> fs_inst *inst)
>     const fs_reg dst = horiz_offset(inst->dst, lbld.group());
>     const unsigned dst_size = inst->regs_written * REG_SIZE /
>              inst->dst.component_size(inst->exec_size);
> +
> +   if (needs_dst_copy(lbld, inst)) {
>     const fs_reg tmp = lbld.vgrf(inst->dst.type, dst_size);
>
>     if (inst->predicate) {
> @@ -5090,6 +5164,13 @@ emit_zip(const fs_builder &lbld, bblock_t *block,
> fs_inst *inst)
>            .MOV(offset(dst, inst->exec_size, k), offset(tmp, lbld, k));
>
>     return tmp;
> +
> +   } else {
> +      /* No need to allocate a temporary for the lowered instruction, just
> +       * take the right group of channels from the original region.
> +       */
> +      return dst;
> +   }
>  }
>
>  bool
> --
> 2.7.3
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160528/b0ee978b/attachment-0001.html>