[Mesa-dev] [PATCH 04/10] nir: Introduce a nir_opt_move_comparisons() pass.
Timothy Arceri
t_arceri at yahoo.com.au
Mon Jan 9 00:13:53 UTC 2017
On Sat, 2017-01-07 at 07:45 -0800, Jason Ekstrand wrote:
> Looks functionally correct. I left a few simple comments.
>
> On Jan 7, 2017 3:59 AM, "Timothy Arceri" <timothy.arceri at collabora.co
> m> wrote:
> From: Kenneth Graunke <kenneth at whitecape.org>
>
> This tries to move comparisons (a common source of boolean values)
> closer to their first use. For GPUs which use condition codes,
> this can eliminate a lot of temporary booleans and comparisons
> which reload the condition code register based on a boolean.
>
> V2: (Timothy Arceri)
> - fix move comparision for phis so we dont end up with:
>
> vec1 32 ssa_227 = phi block_34: ssa_1, block_38: ssa_240
> vec1 32 ssa_235 = feq ssa_227, ssa_1
> vec1 32 ssa_230 = phi block_34: ssa_221, block_38: ssa_235
>
> - add nir_op_i2b/nir_op_f2b to the list of comparisons.
>
> Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> Reviewed-by: Ian Romanick <ian.d.romanick at intel.com> [v1]
> ---
> src/compiler/Makefile.sources | 1 +
> src/compiler/nir/nir.h | 2 +
> src/compiler/nir/nir_opt_move_comparisons.c | 176
> ++++++++++++++++++++++++++++
> 3 files changed, 179 insertions(+)
> create mode 100644 src/compiler/nir/nir_opt_move_comparisons.c
>
> diff --git a/src/compiler/Makefile.sources
> b/src/compiler/Makefile.sources
> index 52f6e54..6da854e 100644
> --- a/src/compiler/Makefile.sources
> +++ b/src/compiler/Makefile.sources
> @@ -245,6 +245,7 @@ NIR_FILES = \
> nir/nir_opt_global_to_local.c \
> nir/nir_opt_if.c \
> nir/nir_opt_loop_unroll.c \
> + nir/nir_opt_move_comparisons.c \
> nir/nir_opt_peephole_select.c \
> nir/nir_opt_remove_phis.c \
> nir/nir_opt_trivial_continues.c \
> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
> index d17924c..325d73b 100644
> --- a/src/compiler/nir/nir.h
> +++ b/src/compiler/nir/nir.h
> @@ -2563,6 +2563,8 @@ bool nir_opt_if(nir_shader *shader);
>
> bool nir_opt_loop_unroll(nir_shader *shader, nir_variable_mode
> indirect_mask);
>
> +bool nir_opt_move_comparisons(nir_shader *shader);
> +
> bool nir_opt_peephole_select(nir_shader *shader, unsigned limit);
>
> bool nir_opt_remove_phis(nir_shader *shader);
> diff --git a/src/compiler/nir/nir_opt_move_comparisons.c
> b/src/compiler/nir/nir_opt_move_comparisons.c
> new file mode 100644
> index 0000000..e6bee94
> --- /dev/null
> +++ b/src/compiler/nir/nir_opt_move_comparisons.c
> @@ -0,0 +1,176 @@
> +/*
> + * Copyright © 2016 Intel Corporation
> + *
> + * Permission is hereby granted, free of charge, to any person
> obtaining a
> + * copy of this software and associated documentation files (the
> "Software"),
> + * to deal in the Software without restriction, including without
> limitation
> + * the rights to use, copy, modify, merge, publish, distribute,
> sublicense,
> + * and/or sell copies of the Software, and to permit persons to whom
> the
> + * Software is furnished to do so, subject to the following
> conditions:
> + *
> + * The above copyright notice and this permission notice (including
> the next
> + * paragraph) shall be included in all copies or substantial
> portions of the
> + * Software.
> + *
> + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
> EXPRESS OR
> + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
> MERCHANTABILITY,
> + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO
> EVENT SHALL
> + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES
> OR OTHER
> + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
> ARISING
> + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
> OTHER DEALINGS
> + * IN THE SOFTWARE.
> + */
> +
> +#include "nir.h"
> +
> +/**
> + * \file nir_opt_move_comparisons.c
> + *
> + * This pass moves ALU comparison operations just before their first
> use.
> + *
> + * It only moves instructions within a single basic block; cross-
> block
> + * movement is left to global code motion.
> + *
> + * Many GPUs generate condition codes for comparisons, and use
> predication
> + * for conditional selects and control flow. In a sequence such as:
> + *
> + * vec1 32 ssa_1 = flt a b
> + * <some other operations>
> + * vec1 32 ssa_2 = bcsel ssa_1 c d
> + *
> + * the backend would likely do the comparison, producing condition
> codes,
> + * then save those to a boolean value. The intervening operations
> might
> + * trash the condition codes. Then, in order to do the bcsel, it
> would
> + * need to re-populate the condition code register based on the
> boolean.
> + *
> + * By moving the comparison just before the bcsel, the condition
> codes could
> + * be used directly. This eliminates the need to reload them from
> the boolean
> + * (generally eliminating an instruction). It may also eliminate
> the need to
> + * create a boolean value altogether (unless it's used elsewhere),
> which could
> + * lower register pressure.
> + */
> +
> +static bool
> +is_comparison(nir_op op)
> +{
> + switch (op) {
> + case nir_op_flt:
> + case nir_op_fge:
> + case nir_op_feq:
> + case nir_op_fne:
> + case nir_op_ilt:
> + case nir_op_ult:
> + case nir_op_ige:
> + case nir_op_uge:
> + case nir_op_ieq:
> + case nir_op_ine:
> + case nir_op_i2b:
> + case nir_op_f2b:
> + return true;
> + default:
> + return false;
> + }
> +}
> +
> +static bool
> +move_comparison_source(nir_src *src, nir_block *block, nir_instr
> *before)
> +{
> + if (src->is_ssa && src->ssa->parent_instr->block == block &&
> + src->ssa->parent_instr->type == nir_instr_type_alu &&
> + is_comparison(nir_instr_as_alu(src->ssa->parent_instr)->op)
> &&
> + (!before || before->type != nir_instr_type_phi)) {
> +
> + struct exec_node *src_node = &src->ssa->parent_instr->node;
>
> I think I would mildly prefer this to be "nir_instr *src_instr". It
> means a bit more typing but the thing you are really doing is
> adding/removing an instruction not a node.
I've made the other two changes and sent a v3. Unfortunately I won't
have much time to work on this stuff any more so would be good if we
could land as is if its only a mild preference.
>
> + exec_node_remove(src_node);
> +
> + if (before)
> + exec_node_insert_node_before(&before->node, src_node);
> + else
> + exec_list_push_tail(&block->instr_list, src_node);
> +
> + return true;
> + }
> +
> + return false;
> +}
> +
> +/* nir_foreach_src callback boilerplate */
> +struct nomc_tuple
> +{
> + nir_instr *instr;
> + bool progress;
> +};
>
> nir_src has a parent_instr field so you can pull everything other
> then the progress boolean out of the SRC and you don't need this
> struct.
>
> +
> +static bool
> +move_comparison_source_cb(nir_src *src, void *data)
> +{
> + struct nomc_tuple *tuple = data;
> +
> + if (move_comparison_source(src, tuple->instr->block, tuple-
> >instr))
> + tuple->progress = true;
> +
> + return true; /* nir_foreach_src should keep going */
> +}
> +
> +static bool
> +move_comparisons(nir_block *block)
> +{
> + bool progress = false;
> +
> + /* We use a simple approach: walk instructions backwards.
> + *
> + * If the instruction's source is a comparison from the same
> block,
> + * simply move it here. This may break SSA if it's used earlier
> in
> + * the block as well. However, as we walk backwards, we'll find
> the
> + * earlier use and move it again, further up. It eventually ends
> up
> + * dominating all uses again, restoring SSA form.
> + *
> + * Before walking instructions, we consider the if-condition at
> the
> + * end of the block, if one exists. It's effectively a use at
> the
> + * bottom of the block.
> + */
> + nir_if *iff = nir_block_get_following_if(block);
> + if (iff) {
> + progress |= move_comparison_source(&iff->condition, block,
> NULL);
> + }
> +
> + nir_foreach_instr_reverse(instr, block) {
> + if (instr->type == nir_instr_type_alu) {
> + /* Walk ALU instruction sources backwards so that bcsel's
> boolean
> + * condition is processed last.
> + */
> + nir_alu_instr *alu = nir_instr_as_alu(instr);
> + for (int i = nir_op_infos[alu->op].num_inputs - 1; i >= 0;
> i--) {
> + progress |= move_comparison_source(&alu->src[i].src,
> + block, instr);
> + }
> + } else {
> + struct nomc_tuple tuple = { instr, false };
> + nir_foreach_src(instr, move_comparison_source_cb, &tuple);
> + progress |= tuple.progress;
> + }
> + }
> +
> + return progress;
> +}
> +
> +bool
> +nir_opt_move_comparisons(nir_shader *shader)
> +{
> + bool progress = false;
> +
> + nir_foreach_function(func, shader) {
> + if (func->impl) {
>
> I tend to prefer "if (!func->impl) continue;" myself but it doesn't
> really matter.
>
> + nir_foreach_block(block, func->impl) {
> + if (move_comparisons(block)) {
> + nir_metadata_preserve(func->impl,
> nir_metadata_block_index |
> +
> nir_metadata_dominance |
> +
> nir_metadata_live_ssa_defs);
> + progress = true;
> + }
> + }
> + }
> + }
> +
> + return progress;
> +}
> --
> 2.9.3
>
> _______________________________________________
> 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