From: Iago Toral Quiroga Date: Wed, 24 Oct 2018 07:25:29 +0000 (+0200) Subject: nir: add a bool bitsize lowering pass X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=467c9a0faa0a27a14bbf5fb42bbf806382d2d3f0;p=mesa.git nir: add a bool bitsize lowering pass The pass lowers 1-bit booleans produced by NIR to the native bitsize of the operations that produce them. v2: change on lower_load_const_instr after upstream changes. Added TODO2 to explain it, as it was not properly tested yet (see already existing TODO) (Neil) Reviewed-by: Alyssa Rosenzweig Reviewed-by: Kristian H. Kristensen Part-of: --- diff --git a/src/compiler/Makefile.sources b/src/compiler/Makefile.sources index 1b05b6bd955..170201e3d34 100644 --- a/src/compiler/Makefile.sources +++ b/src/compiler/Makefile.sources @@ -239,6 +239,7 @@ 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_bitsize.c \ nir/nir_lower_bool_to_float.c \ nir/nir_lower_bool_to_int32.c \ nir/nir_lower_clamp_color_outputs.c \ diff --git a/src/compiler/nir/meson.build b/src/compiler/nir/meson.build index 37c64cab9fd..3e48eb1cf6a 100644 --- a/src/compiler/nir/meson.build +++ b/src/compiler/nir/meson.build @@ -118,6 +118,7 @@ files_libnir = files( 'nir_lower_array_deref_of_vec.c', 'nir_lower_atomics_to_ssbo.c', 'nir_lower_bitmap.c', + 'nir_lower_bool_to_bitsize.c', 'nir_lower_bool_to_float.c', 'nir_lower_bool_to_int32.c', 'nir_lower_clamp_color_outputs.c', diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h index 1fe0264cbe5..6a1a8e175df 100644 --- a/src/compiler/nir/nir.h +++ b/src/compiler/nir/nir.h @@ -3868,6 +3868,7 @@ bool nir_lower_flrp(nir_shader *shader, unsigned lowering_mask, bool always_precise, bool have_ffma); bool nir_lower_alu_to_scalar(nir_shader *shader, nir_instr_filter_cb cb, const void *data); +bool nir_lower_bool_to_bitsize(nir_shader *shader); bool nir_lower_bool_to_float(nir_shader *shader); bool nir_lower_bool_to_int32(nir_shader *shader); bool nir_lower_int_to_float(nir_shader *shader); diff --git a/src/compiler/nir/nir_lower_bool_to_bitsize.c b/src/compiler/nir/nir_lower_bool_to_bitsize.c new file mode 100644 index 00000000000..5561ffa14a4 --- /dev/null +++ b/src/compiler/nir/nir_lower_bool_to_bitsize.c @@ -0,0 +1,431 @@ +/* + * 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 uint32_t +get_bool_convert_opcode(uint32_t dst_bit_size) +{ + switch (dst_bit_size) { + case 32: return nir_op_i2i32; + case 16: return nir_op_i2i16; + case 8: return nir_op_i2i8; + default: + unreachable("invalid boolean bit-size"); + } +} + +static void +make_sources_canonical(nir_builder *b, nir_alu_instr *alu, uint32_t start_idx) +{ + /* TODO: for now we take the bit-size of the first source as the canonical + * form but we could try to be smarter. + */ + const nir_op_info *op_info = &nir_op_infos[alu->op]; + uint32_t bit_size = nir_src_bit_size(alu->src[start_idx].src); + for (uint32_t i = start_idx + 1; i < op_info->num_inputs; i++) { + if (nir_src_bit_size(alu->src[i].src) != bit_size) { + b->cursor = nir_before_instr(&alu->instr); + nir_op convert_op = get_bool_convert_opcode(bit_size); + nir_ssa_def *new_src = + nir_build_alu(b, convert_op, alu->src[i].src.ssa, NULL, NULL, NULL); + /* Retain the write mask and swizzle of the original instruction so + * that we don’t unnecessarily create a vectorized instruction. + */ + nir_alu_instr *conv_instr = + nir_instr_as_alu(nir_builder_last_instr(b)); + conv_instr->dest.write_mask = alu->dest.write_mask; + conv_instr->dest.dest.ssa.num_components = + alu->dest.dest.ssa.num_components; + memcpy(conv_instr->src[0].swizzle, + alu->src[i].swizzle, + sizeof(conv_instr->src[0].swizzle)); + nir_instr_rewrite_src(&alu->instr, + &alu->src[i].src, nir_src_for_ssa(new_src)); + /* The swizzle will have been handled by the conversion instruction + * so we can reset it back to the default + */ + for (unsigned j = 0; j < NIR_MAX_VEC_COMPONENTS; j++) + alu->src[i].swizzle[j] = j; + } + } +} + +static bool +lower_alu_instr(nir_builder *b, nir_alu_instr *alu) +{ + const nir_op_info *op_info = &nir_op_infos[alu->op]; + + /* For operations that can take multiple boolean sources we need to ensure + * that all booleans have the same bit-size + */ + switch (alu->op) { + case nir_op_mov: + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + case nir_op_inot: + case nir_op_iand: + case nir_op_ior: + case nir_op_ixor: + if (nir_dest_bit_size(alu->dest.dest) > 1) + break; /* Not a boolean instruction */ + /* Fallthrough */ + + case nir_op_ball_fequal2: + case nir_op_ball_fequal3: + case nir_op_ball_fequal4: + case nir_op_bany_fnequal2: + case nir_op_bany_fnequal3: + case nir_op_bany_fnequal4: + case nir_op_ball_iequal2: + case nir_op_ball_iequal3: + case nir_op_ball_iequal4: + case nir_op_bany_inequal2: + case nir_op_bany_inequal3: + case nir_op_bany_inequal4: + case nir_op_ieq: + case nir_op_ine: + make_sources_canonical(b, alu, 0); + break; + + case nir_op_bcsel: + /* bcsel may be choosing between boolean sources too */ + if (nir_dest_bit_size(alu->dest.dest) == 1) + make_sources_canonical(b, alu, 1); + break; + + default: + break; + } + + /* Now that we have a canonical boolean bit-size, go on and rewrite the + * instruction to match the canonical bit-size. + */ + uint32_t bit_size = nir_src_bit_size(alu->src[0].src); + assert(bit_size > 1); + + nir_op opcode = alu->op; + switch (opcode) { + case nir_op_mov: + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + case nir_op_inot: + case nir_op_iand: + case nir_op_ior: + case nir_op_ixor: + /* Nothing to do here, we do not specialize these opcodes by bit-size */ + break; + + case nir_op_f2b1: + opcode = bit_size == 8 ? nir_op_f2b8 : + bit_size == 16 ? nir_op_f2b16 : nir_op_f2b32; + break; + + case nir_op_i2b1: + opcode = bit_size == 8 ? nir_op_i2b8 : + bit_size == 16 ? nir_op_i2b16 : nir_op_i2b32; + break; + + case nir_op_flt: + opcode = bit_size == 8 ? nir_op_flt8 : + bit_size == 16 ? nir_op_flt16 : nir_op_flt32; + break; + + case nir_op_fge: + opcode = bit_size == 8 ? nir_op_fge8 : + bit_size == 16 ? nir_op_fge16 : nir_op_fge32; + break; + + case nir_op_feq: + opcode = bit_size == 8 ? nir_op_feq8 : + bit_size == 16 ? nir_op_feq16 : nir_op_feq32; + break; + + case nir_op_fne: + opcode = bit_size == 8 ? nir_op_fne8 : + bit_size == 16 ? nir_op_fne16 : nir_op_fne32; + break; + + case nir_op_ilt: + opcode = bit_size == 8 ? nir_op_ilt8 : + bit_size == 16 ? nir_op_ilt16 : nir_op_ilt32; + break; + + case nir_op_ige: + opcode = bit_size == 8 ? nir_op_ige8 : + bit_size == 16 ? nir_op_ige16 : nir_op_ige32; + break; + + case nir_op_ieq: + opcode = bit_size == 8 ? nir_op_ieq8 : + bit_size == 16 ? nir_op_ieq16 : nir_op_ieq32; + break; + + case nir_op_ine: + opcode = bit_size == 8 ? nir_op_ine8 : + bit_size == 16 ? nir_op_ine16 : nir_op_ine32; + break; + + case nir_op_ult: + opcode = bit_size == 8 ? nir_op_ult8 : + bit_size == 16 ? nir_op_ult16 : nir_op_ult32; + break; + + case nir_op_uge: + opcode = bit_size == 8 ? nir_op_uge8 : + bit_size == 16 ? nir_op_uge16 : nir_op_uge32; + break; + + case nir_op_ball_fequal2: + opcode = bit_size == 8 ? nir_op_b8all_fequal2 : + bit_size == 16 ? nir_op_b16all_fequal2 : + nir_op_b32all_fequal2; + break; + + case nir_op_ball_fequal3: + opcode = bit_size == 8 ? nir_op_b8all_fequal3 : + bit_size == 16 ? nir_op_b16all_fequal3 : + nir_op_b32all_fequal3; + break; + + case nir_op_ball_fequal4: + opcode = bit_size == 8 ? nir_op_b8all_fequal4 : + bit_size == 16 ? nir_op_b16all_fequal4 : + nir_op_b32all_fequal4; + break; + + case nir_op_bany_fnequal2: + opcode = bit_size == 8 ? nir_op_b8any_fnequal2 : + bit_size == 16 ? nir_op_b16any_fnequal2 : + nir_op_b32any_fnequal2; + break; + + case nir_op_bany_fnequal3: + opcode = bit_size == 8 ? nir_op_b8any_fnequal3 : + bit_size == 16 ? nir_op_b16any_fnequal3 : + nir_op_b32any_fnequal3; + break; + + case nir_op_bany_fnequal4: + opcode = bit_size == 8 ? nir_op_b8any_fnequal4 : + bit_size == 16 ? nir_op_b16any_fnequal4 : + nir_op_b32any_fnequal4; + break; + + case nir_op_ball_iequal2: + opcode = bit_size == 8 ? nir_op_b8all_iequal2 : + bit_size == 16 ? nir_op_b16all_iequal2 : + nir_op_b32all_iequal2; + break; + + case nir_op_ball_iequal3: + opcode = bit_size == 8 ? nir_op_b8all_iequal3 : + bit_size == 16 ? nir_op_b16all_iequal3 : + nir_op_b32all_iequal3; + break; + + case nir_op_ball_iequal4: + opcode = bit_size == 8 ? nir_op_b8all_iequal4 : + bit_size == 16 ? nir_op_b16all_iequal4 : + nir_op_b32all_iequal4; + break; + + case nir_op_bany_inequal2: + opcode = bit_size == 8 ? nir_op_b8any_inequal2 : + bit_size == 16 ? nir_op_b16any_inequal2 : + nir_op_b32any_inequal2; + break; + + case nir_op_bany_inequal3: + opcode = bit_size == 8 ? nir_op_b8any_inequal3 : + bit_size == 16 ? nir_op_b16any_inequal3 : + nir_op_b32any_inequal3; + break; + + case nir_op_bany_inequal4: + opcode = bit_size == 8 ? nir_op_b8any_inequal4 : + bit_size == 16 ? nir_op_b16any_inequal4 : + nir_op_b32any_inequal4; + break; + + case nir_op_bcsel: + opcode = bit_size == 8 ? nir_op_b8csel : + bit_size == 16 ? nir_op_b16csel : nir_op_b32csel; + + /* The destination of the selection may have a different bit-size from + * the bcsel condition. + */ + bit_size = nir_src_bit_size(alu->src[1].src); + break; + + 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; + } + + alu->op = opcode; + + if (alu->dest.dest.ssa.bit_size == 1) + alu->dest.dest.ssa.bit_size = bit_size; + + return true; +} + +static bool +lower_load_const_instr(nir_load_const_instr *load) +{ + bool progress = false; + + if (load->def.bit_size > 1) + return progress; + + /* TODO: It is not clear if there is any case in which we can ever hit + * this path, so for now we just provide a 32-bit default. + * + * TODO2: after some changed on nir_const_value and other on upstream, we + * removed the initialization of a general value like this: + * nir_const_value value = load->value + * + * to initialize per value component. Need to confirm if that is correct, + * but look at the TOO before. + */ + for (unsigned i = 0; i < load->def.num_components; i++) { + load->value[i].u32 = load->value[i].b ? NIR_TRUE : NIR_FALSE; + load->def.bit_size = 32; + progress = true; + } + + return progress; +} + +static bool +lower_phi_instr(nir_builder *b, nir_phi_instr *phi) +{ + if (nir_dest_bit_size(phi->dest) != 1) + return false; + + /* Ensure all phi sources have a canonical bit-size. We choose the + * bit-size of the first phi source as the canonical form. + * + * TODO: maybe we can be smarter about how we choose the canonical form. + */ + uint32_t dst_bit_size = 0; + nir_foreach_phi_src(phi_src, phi) { + uint32_t src_bit_size = nir_src_bit_size(phi_src->src); + if (dst_bit_size == 0) { + dst_bit_size = src_bit_size; + } else if (src_bit_size != dst_bit_size) { + assert(phi_src->src.is_ssa); + b->cursor = nir_before_src(&phi_src->src, false); + nir_op convert_op = get_bool_convert_opcode(dst_bit_size); + nir_ssa_def *new_src = + nir_build_alu(b, convert_op, phi_src->src.ssa, NULL, NULL, NULL); + nir_instr_rewrite_src(&phi->instr, &phi_src->src, + nir_src_for_ssa(new_src)); + } + } + + phi->dest.ssa.bit_size = dst_bit_size; + + return true; +} + +static bool +nir_lower_bool_to_bitsize_impl(nir_builder *b, nir_function_impl *impl) +{ + bool progress = false; + + 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: + progress |= lower_load_const_instr(nir_instr_as_load_const(instr)); + break; + + case nir_instr_type_phi: + progress |= lower_phi_instr(b, nir_instr_as_phi(instr)); + break; + + case nir_instr_type_ssa_undef: + case nir_instr_type_intrinsic: + 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_bitsize(nir_shader *shader) +{ + nir_builder b; + bool progress = false; + + nir_foreach_function(function, shader) { + if (function->impl) { + nir_builder_init(&b, function->impl); + progress = nir_lower_bool_to_bitsize_impl(&b, function->impl) || progress; + } + } + + return progress; +}