[Mesa-dev] [RFC 31/31] nir: Add a bool to float32 lowering pass
Ian Romanick
idr at freedesktop.org
Tue Oct 23 16:31:18 UTC 2018
On 10/23/2018 08:33 AM, Connor Abbott wrote:
> On Tue, Oct 23, 2018 at 12:16 AM Jason Ekstrand <jason at jlekstrand.net> wrote:
>>
>> This should be useful for drivers that don't support real integers.
>>
>> Cc: Alyssa Rosenzweig <alyssa at rosenzweig.io>
>> ---
>> src/compiler/Makefile.sources | 1 +
>> src/compiler/nir/meson.build | 1 +
>> src/compiler/nir/nir_lower_bool_to_float.c | 181 +++++++++++++++++++++
>> 3 files changed, 183 insertions(+)
>> create mode 100644 src/compiler/nir/nir_lower_bool_to_float.c
>>
>> diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources
>> index 8f65f974ab8..2ff12ff43cb 100644
>> --- a/src/compiler/Makefile.sources
>> +++ b/src/compiler/Makefile.sources
>> @@ -230,6 +230,7 @@ NIR_FILES = \
>> nir/nir_lower_atomics_to_ssbo.c \
>> nir/nir_lower_bitmap.c \
>> nir/nir_lower_bit_size.c \
>> + nir/nir_lower_bool_to_float.c \
>> nir/nir_lower_bool_to_int32.c \
>> nir/nir_lower_clamp_color_outputs.c \
>> nir/nir_lower_clip.c \
>> diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build
>> index 5809551c9d4..f715668a03b 100644
>> --- a/src/compiler/nir/meson.build
>> +++ b/src/compiler/nir/meson.build
>> @@ -113,6 +113,7 @@ files_libnir = files(
>> 'nir_lower_alpha_test.c',
>> 'nir_lower_atomics_to_ssbo.c',
>> 'nir_lower_bitmap.c',
>> + 'nir_lower_bool_to_float.c',
>> 'nir_lower_bool_to_int32.c',
>> 'nir_lower_clamp_color_outputs.c',
>> 'nir_lower_clip.c',
>> diff --git a/src/compiler/nir/nir_lower_bool_to_float.c b/src/compiler/nir/nir_lower_bool_to_float.c
>> new file mode 100644
>> index 00000000000..7aa5efb5a2f
>> --- /dev/null
>> +++ b/src/compiler/nir/nir_lower_bool_to_float.c
>> @@ -0,0 +1,181 @@
>> +/*
>> + * Copyright © 2018 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"
>> +#include "nir_builder.h"
>> +
>> +static bool
>> +assert_ssa_def_is_not_1bit(nir_ssa_def *def, UNUSED void *unused)
>> +{
>> + assert(def->bit_size > 1);
>> + return true;
>> +}
>> +
>> +static bool
>> +rewrite_1bit_ssa_def_to_32bit(nir_ssa_def *def, void *_progress)
>> +{
>> + bool *progress = _progress;
>> + if (def->bit_size == 1) {
>> + def->bit_size = 32;
>> + *progress = true;
>> + }
>> + return true;
>> +}
>> +
>> +static bool
>> +lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
>> +{
>> + const nir_op_info *op_info = &nir_op_infos[alu->op];
>> +
>> + b->cursor = nir_before_instr(&alu->instr);
>> +
>> + /* Replacement SSA value */
>> + nir_ssa_def *rep = NULL;
>> + switch (alu->op) {
>> + case nir_op_b2f: alu->op = nir_op_fmov; break;
>> + case nir_op_b2i: alu->op = nir_op_fmov; break;
>> + case nir_op_f2b:
>> + case nir_op_i2b:
>> + rep = nir_sne(b, nir_ssa_for_alu_src(b, alu, 0),
>> + nir_imm_float(b, 0));
>> + break;
>> +
>> + case nir_op_flt: alu->op = nir_op_slt; break;
>> + case nir_op_fge: alu->op = nir_op_sge; break;
>> + case nir_op_feq: alu->op = nir_op_seq; break;
>> + case nir_op_fne: alu->op = nir_op_sne; break;
>> + case nir_op_ilt: alu->op = nir_op_slt; break;
>> + case nir_op_ige: alu->op = nir_op_sge; break;
>> + case nir_op_ieq: alu->op = nir_op_seq; break;
>> + case nir_op_ine: alu->op = nir_op_sne; break;
>> + case nir_op_ult: alu->op = nir_op_slt; break;
>> + case nir_op_uge: alu->op = nir_op_sge; break;
>> +
>> + case nir_op_ball_fequal2: alu->op = nir_op_fall_equal2; break;
>> + case nir_op_ball_fequal3: alu->op = nir_op_fall_equal3; break;
>> + case nir_op_ball_fequal4: alu->op = nir_op_fall_equal4; break;
>> + case nir_op_bany_fnequal2: alu->op = nir_op_fany_nequal2; break;
>> + case nir_op_bany_fnequal3: alu->op = nir_op_fany_nequal3; break;
>> + case nir_op_bany_fnequal4: alu->op = nir_op_fany_nequal4; break;
>> + case nir_op_ball_iequal2: alu->op = nir_op_fall_equal2; break;
>> + case nir_op_ball_iequal3: alu->op = nir_op_fall_equal3; break;
>> + case nir_op_ball_iequal4: alu->op = nir_op_fall_equal4; break;
>> + case nir_op_bany_inequal2: alu->op = nir_op_fany_nequal2; break;
>> + case nir_op_bany_inequal3: alu->op = nir_op_fany_nequal3; break;
>> + case nir_op_bany_inequal4: alu->op = nir_op_fany_nequal4; break;
>> +
>> + case nir_op_bcsel: alu->op = nir_op_fcsel; break;
>> +
>> + case nir_op_imov: alu->op = nir_op_fmov; break;
>> + case nir_op_iand: alu->op = nir_op_fmul; break;
>> + case nir_op_ixor: alu->op = nir_op_sne; break;
>> +
>> + case nir_op_ior:
>> + rep = nir_fsat(b, nir_fadd(b, nir_ssa_for_alu_src(b, alu, 0),
>> + nir_ssa_for_alu_src(b, alu, 1)));
>
> At least on old Mali, this would be faster as fmax since fsat is an
> extra instruction.
I guess it will depend on what other non-integer GPUs can do, but we may
want a flag for this. On GPUs where .sat is free, using add for or
means that (a && b) || c can become ffma.sat.
>> + break;
>> + case nir_op_inot:
>> + rep = nir_fsub(b, nir_imm_float(b, 1.0),
>> + nir_ssa_for_alu_src(b, alu, 0));
>> + break;
>
> Not sure if it matters, but apparently we already have an fnot
> instruction. I don't know if any hw can do something better than this,
> though.
>
>> +
>> + default:
>> + assert(alu->dest.dest.ssa.bit_size > 1);
>> + for (unsigned i = 0; i < op_info->num_inputs; i++)
>> + assert(alu->src[i].src.ssa->bit_size > 1);
>> + return false;
>> + }
>> +
>> + if (rep) {
>> + /* We've emitted a replacement instruction */
>> + nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(rep));
>> + nir_instr_remove(&alu->instr);
>> + } else {
>> + if (alu->dest.dest.ssa.bit_size == 1)
>> + alu->dest.dest.ssa.bit_size = 32;
>> + }
>> +
>> + return true;
>> +}
>> +
>> +static bool
>> +nir_lower_bool_to_int32_impl(nir_function_impl *impl)
>
> Whoops, int32...
>
>> +{
>> + bool progress = false;
>> +
>> + nir_builder b;
>> + nir_builder_init(&b, impl);
>> +
>> + nir_foreach_block(block, impl) {
>> + nir_foreach_instr_safe(instr, block) {
>> + switch (instr->type) {
>> + case nir_instr_type_alu:
>> + progress |= lower_alu_instr(&b, nir_instr_as_alu(instr));
>> + break;
>> +
>> + case nir_instr_type_load_const: {
>> + nir_load_const_instr *load = nir_instr_as_load_const(instr);
>> + if (load->def.bit_size == 1) {
>> + nir_const_value value = load->value;
>> + for (unsigned i = 0; i < load->def.num_components; i++)
>> + load->value.u32[i] = value.b[i] ? NIR_TRUE : NIR_FALSE;
>
> value.f32[i] = value.b[i] ? 1.0 : 0.0
>
>> + load->def.bit_size = 32;
>> + progress = true;
>> + }
>> + break;
>> + }
>> +
>> + case nir_instr_type_intrinsic:
>> + case nir_instr_type_ssa_undef:
>> + case nir_instr_type_phi:
>> + case nir_instr_type_tex:
>> + nir_foreach_ssa_def(instr, rewrite_1bit_ssa_def_to_32bit,
>> + &progress);
>> + break;
>> +
>> + default:
>> + nir_foreach_ssa_def(instr, assert_ssa_def_is_not_1bit, NULL);
>> + }
>> + }
>> + }
>> +
>> + if (progress) {
>> + nir_metadata_preserve(impl, nir_metadata_block_index |
>> + nir_metadata_dominance);
>> + }
>> +
>> + return progress;
>> +}
>> +
>> +bool
>> +nir_lower_bool_to_int32(nir_shader *shader)
>
> int32...
>
>> +{
>> + bool progress = false;
>> +
>> + nir_foreach_function(function, shader) {
>> + if (function->impl && nir_lower_bool_to_int32_impl(function->impl))
>> + progress = true;
>> + }
>> +
>> + return progress;
>> +}
>> --
>> 2.19.1
>>
>> _______________________________________________
>> 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