mirror of
https://gitlab.freedesktop.org/mesa/mesa.git
synced 2025-12-22 22:10:10 +01:00
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io> Acked-by: Faith Ekstrand <faith.ekstrand@collabora.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33722>
344 lines
11 KiB
C
344 lines
11 KiB
C
/*
|
|
* Copyright © 2015 Broadcom
|
|
*
|
|
* 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 <math.h>
|
|
#include "util/mesa-blake3.h"
|
|
#include "nir.h"
|
|
#include "nir_builder.h"
|
|
|
|
/** @file nir_opt_undef.c
|
|
*
|
|
* Handles optimization of operations involving ssa_undef.
|
|
*/
|
|
|
|
struct undef_options {
|
|
bool disallow_undef_to_nan;
|
|
};
|
|
|
|
/**
|
|
* Turn conditional selects between an undef and some other value into a move
|
|
* of that other value (on the assumption that the condition's going to be
|
|
* choosing the defined value). This reduces work after if flattening when
|
|
* each side of the if is defining a variable.
|
|
*/
|
|
static bool
|
|
opt_undef_csel(nir_builder *b, nir_alu_instr *instr)
|
|
{
|
|
if (!nir_op_is_selection(instr->op))
|
|
return false;
|
|
|
|
for (int i = 1; i <= 2; i++) {
|
|
if (!nir_src_is_undef(instr->src[i].src))
|
|
continue;
|
|
|
|
b->cursor = nir_instr_remove(&instr->instr);
|
|
nir_def *mov = nir_mov_alu(b, instr->src[i == 1 ? 2 : 1],
|
|
instr->def.num_components);
|
|
nir_def_rewrite_uses(&instr->def, mov);
|
|
|
|
return true;
|
|
}
|
|
|
|
return false;
|
|
}
|
|
|
|
static bool
|
|
op_is_mov_or_vec_or_pack_or_unpack(nir_op op)
|
|
{
|
|
switch (op) {
|
|
case nir_op_pack_32_2x16:
|
|
case nir_op_pack_32_2x16_split:
|
|
case nir_op_pack_32_4x8:
|
|
case nir_op_pack_32_4x8_split:
|
|
case nir_op_pack_64_2x32:
|
|
case nir_op_pack_64_2x32_split:
|
|
case nir_op_pack_64_4x16:
|
|
case nir_op_unpack_32_2x16:
|
|
case nir_op_unpack_32_2x16_split_x:
|
|
case nir_op_unpack_32_2x16_split_y:
|
|
case nir_op_unpack_32_4x8:
|
|
case nir_op_unpack_64_2x32:
|
|
case nir_op_unpack_64_2x32_split_x:
|
|
case nir_op_unpack_64_2x32_split_y:
|
|
case nir_op_unpack_64_4x16:
|
|
return true;
|
|
default:
|
|
return nir_op_is_vec_or_mov(op);
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Replace vecN(undef, undef, ...) with a single undef.
|
|
*/
|
|
static bool
|
|
opt_undef_vecN(nir_builder *b, nir_alu_instr *alu)
|
|
{
|
|
if (!op_is_mov_or_vec_or_pack_or_unpack(alu->op))
|
|
return false;
|
|
|
|
for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
|
|
if (!nir_src_is_undef(alu->src[i].src))
|
|
return false;
|
|
}
|
|
|
|
b->cursor = nir_before_instr(&alu->instr);
|
|
nir_def *undef = nir_undef(b, alu->def.num_components,
|
|
alu->def.bit_size);
|
|
nir_def_replace(&alu->def, undef);
|
|
|
|
return true;
|
|
}
|
|
|
|
static uint32_t
|
|
nir_get_undef_mask(nir_def *def)
|
|
{
|
|
nir_instr *instr = def->parent_instr;
|
|
|
|
if (instr->type == nir_instr_type_undef)
|
|
return BITSET_MASK(def->num_components);
|
|
|
|
if (instr->type != nir_instr_type_alu)
|
|
return 0;
|
|
|
|
nir_alu_instr *alu = nir_instr_as_alu(instr);
|
|
unsigned undef = 0;
|
|
|
|
/* nir_op_mov of undef is handled by opt_undef_vecN() */
|
|
if (nir_op_is_vec(alu->op)) {
|
|
for (int i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
|
|
if (nir_src_is_undef(alu->src[i].src)) {
|
|
undef |= BITSET_MASK(nir_ssa_alu_instr_src_components(alu, i)) << i;
|
|
}
|
|
}
|
|
}
|
|
|
|
return undef;
|
|
}
|
|
|
|
/**
|
|
* Remove any store intrinsic writemask channels whose value is undefined (the
|
|
* existing value is a fine representation of "undefined").
|
|
*/
|
|
static bool
|
|
opt_undef_store(nir_intrinsic_instr *intrin)
|
|
{
|
|
int arg_index;
|
|
switch (intrin->intrinsic) {
|
|
case nir_intrinsic_store_deref:
|
|
arg_index = 1;
|
|
break;
|
|
case nir_intrinsic_store_output:
|
|
case nir_intrinsic_store_per_vertex_output:
|
|
case nir_intrinsic_store_per_view_output:
|
|
case nir_intrinsic_store_per_primitive_output:
|
|
case nir_intrinsic_store_ssbo:
|
|
case nir_intrinsic_store_shared:
|
|
case nir_intrinsic_store_global:
|
|
case nir_intrinsic_store_scratch:
|
|
arg_index = 0;
|
|
break;
|
|
default:
|
|
return false;
|
|
}
|
|
|
|
nir_def *def = intrin->src[arg_index].ssa;
|
|
|
|
unsigned write_mask = nir_intrinsic_write_mask(intrin);
|
|
unsigned undef_mask = nir_get_undef_mask(def);
|
|
|
|
if (!(write_mask & undef_mask))
|
|
return false;
|
|
|
|
write_mask &= ~undef_mask;
|
|
if (!write_mask)
|
|
nir_instr_remove(&intrin->instr);
|
|
else
|
|
nir_intrinsic_set_write_mask(intrin, write_mask);
|
|
|
|
return true;
|
|
}
|
|
|
|
struct visit_info {
|
|
bool replace_undef_with_constant;
|
|
bool prefer_nan;
|
|
bool must_keep_undef;
|
|
};
|
|
|
|
/**
|
|
* Analyze an undef use to see if replacing undef with a constant is
|
|
* beneficial.
|
|
*/
|
|
static void
|
|
visit_undef_use(nir_src *src, struct visit_info *info)
|
|
{
|
|
if (nir_src_is_if(src)) {
|
|
/* If the use is "if", keep undef because the branch will be eliminated
|
|
* by nir_opt_dead_cf.
|
|
*/
|
|
info->must_keep_undef = true;
|
|
return;
|
|
}
|
|
|
|
nir_instr *instr = nir_src_parent_instr(src);
|
|
|
|
if (instr->type == nir_instr_type_alu) {
|
|
/* Replacing undef with a constant is only beneficial with ALU
|
|
* instructions because it can eliminate them or simplify them.
|
|
*/
|
|
nir_alu_instr *alu = nir_instr_as_alu(instr);
|
|
|
|
/* opt_undef_vecN already copy propagated. */
|
|
if (op_is_mov_or_vec_or_pack_or_unpack(alu->op)) {
|
|
info->must_keep_undef = true;
|
|
return;
|
|
}
|
|
|
|
unsigned num_srcs = nir_op_infos[alu->op].num_inputs;
|
|
|
|
for (unsigned i = 0; i < num_srcs; i++) {
|
|
if (&alu->src[i].src != src)
|
|
continue;
|
|
|
|
info->replace_undef_with_constant = true;
|
|
if (nir_op_infos[alu->op].input_types[i] & nir_type_float &&
|
|
alu->op != nir_op_fmulz &&
|
|
(alu->op != nir_op_ffmaz || i == 2))
|
|
info->prefer_nan = true;
|
|
}
|
|
} else {
|
|
/* If the use is not ALU, don't replace undef. We need to preserve
|
|
* undef for stores and phis because those are handled differently,
|
|
* and replacing undef with a constant would result in worse code.
|
|
*/
|
|
info->must_keep_undef = true;
|
|
return;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Replace ssa_undef used by ALU opcodes with 0 or NaN, whichever eliminates
|
|
* more code.
|
|
*
|
|
* Replace it with NaN if an FP opcode uses undef, which will cause the opcode
|
|
* to be eliminated by nir_opt_algebraic. 0 would not eliminate the FP opcode.
|
|
*/
|
|
static bool
|
|
replace_ssa_undef(nir_builder *b, nir_instr *instr, void *data)
|
|
{
|
|
if (instr->type != nir_instr_type_undef)
|
|
return false;
|
|
|
|
const struct undef_options *options = data;
|
|
|
|
nir_undef_instr *undef = nir_instr_as_undef(instr);
|
|
struct visit_info info = { 0 };
|
|
|
|
nir_foreach_use_including_if(src, &undef->def) {
|
|
visit_undef_use(src, &info);
|
|
}
|
|
|
|
if (info.must_keep_undef || !info.replace_undef_with_constant)
|
|
return false;
|
|
|
|
b->cursor = nir_before_instr(&undef->instr);
|
|
nir_def *replacement;
|
|
|
|
/* If undef is used as float, replace it with NaN, which will
|
|
* eliminate all FP instructions that consume it. Else, replace it
|
|
* with 0, which is more likely to eliminate non-FP instructions.
|
|
*/
|
|
if (info.prefer_nan && !options->disallow_undef_to_nan)
|
|
replacement = nir_imm_floatN_t(b, NAN, undef->def.bit_size);
|
|
else
|
|
replacement = nir_imm_intN_t(b, 0, undef->def.bit_size);
|
|
|
|
if (undef->def.num_components > 1)
|
|
replacement = nir_replicate(b, replacement, undef->def.num_components);
|
|
|
|
nir_def_replace(&undef->def, replacement);
|
|
return true;
|
|
}
|
|
|
|
static bool
|
|
opt_undef_uses(nir_builder *b, nir_instr *instr, void *data)
|
|
{
|
|
if (instr->type == nir_instr_type_alu) {
|
|
nir_alu_instr *alu = nir_instr_as_alu(instr);
|
|
return opt_undef_csel(b, alu) ||
|
|
opt_undef_vecN(b, alu);
|
|
} else if (instr->type == nir_instr_type_intrinsic) {
|
|
nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
|
|
return opt_undef_store(intrin);
|
|
}
|
|
|
|
return false;
|
|
}
|
|
|
|
bool
|
|
nir_opt_undef(nir_shader *shader)
|
|
{
|
|
struct undef_options options = { 0 };
|
|
|
|
/* Disallow the undef->NaN transformation only for those shaders where
|
|
* it's known to break rendering. These are shader source BLAKE3s printed by
|
|
* nir_print_shader().
|
|
*/
|
|
uint32_t shader_blake3s[][BLAKE3_OUT_LEN32] = {
|
|
/* gputest/gimark */
|
|
{ 0x582c214b, 0x25478275, 0xc9a835d2, 0x95c9b643, 0x69deae47, 0x213c7427, 0xa9da66a5, 0xac254ed2 },
|
|
|
|
/* Viewperf13/CATIA_car_01 */
|
|
{ 0x880dfa0f, 0x60e32201, 0xe3a89f59, 0xb1cc6f07, 0xcdbebe66, 0x20122aec, 0x83450d4e, 0x8f42843d }, /* Taillights */
|
|
{ 0x624e53bb, 0x8eb635ba, 0xb1e4ed9b, 0x651b0fec, 0x86fcf79a, 0xde0863fb, 0x09ce80c1, 0xd972e40f }, /* Grill */
|
|
{ 0x01a8db39, 0xfa175175, 0x621f7302, 0xfcde9177, 0x72d873bf, 0x048d38c1, 0xe669d2de, 0xaa6584af }, /* Headlights */
|
|
{ 0x32029770, 0xab295b41, 0x3f1daf07, 0x9dd9153e, 0xd598be73, 0xe555b2f3, 0x6e087eaf, 0x084d329c }, /* Rims */
|
|
|
|
/* Viewperf13/CATIA_car_04 */
|
|
{ 0x55207b90, 0x08fa2f8f, 0x9db62464, 0xadba6570, 0xb6d5d962, 0xf434bff5, 0x46a34d64, 0x021bfb45 }, /* Headlights */
|
|
{ 0x83fbdd6a, 0x231b027e, 0x6f142248, 0x2b3045de, 0xd2a4f460, 0x59dfb8d8, 0x6dbc00f9, 0xcca13143 }, /* Rims */
|
|
{ 0x88ed3a0a, 0xf128d384, 0x8161fdac, 0xd10cb257, 0x5e63db2d, 0x56798b6f, 0x881e81ee, 0xa4e937d4 }, /* Tires */
|
|
{ 0xbf84697c, 0x3bc75bb6, 0x9d012175, 0x2dd90bcf, 0x0562f0ed, 0x5aa80e62, 0xb5793ae3, 0x9127bcab }, /* Windows */
|
|
{ 0x47a3eb4b, 0x136f676d, 0x94045ed3, 0x57b00972, 0x8cda7550, 0x88327fda, 0x37f7cf37, 0x66db05e3 }, /* Body */
|
|
};
|
|
|
|
for (unsigned i = 0; i < ARRAY_SIZE(shader_blake3s); i++) {
|
|
if (_mesa_printed_blake3_equal(shader->info.source_blake3, shader_blake3s[i])) {
|
|
options.disallow_undef_to_nan = true;
|
|
break;
|
|
}
|
|
}
|
|
|
|
if (shader->info.use_legacy_math_rules)
|
|
options.disallow_undef_to_nan = true;
|
|
|
|
bool progress = nir_shader_instructions_pass(shader,
|
|
opt_undef_uses,
|
|
nir_metadata_control_flow,
|
|
&options);
|
|
progress |= nir_shader_instructions_pass(shader,
|
|
replace_ssa_undef,
|
|
nir_metadata_control_flow,
|
|
&options);
|
|
|
|
return progress;
|
|
}
|