/* * Copyright © 2018 Intel Corporation * Copyright © 2019 Vasily Khoruzhick * * 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_int(nir_def *def, void *arg) { ASSERTED BITSET_WORD *int_types = arg; assert(!BITSET_TEST(int_types, def->index)); return true; } static bool instr_has_only_trivial_swizzles(nir_alu_instr *alu) { const nir_op_info *info = &nir_op_infos[alu->op]; for (unsigned i = 0; i < info->num_inputs; i++) { for (unsigned chan = 0; chan < alu->def.num_components; chan++) { if (alu->src[i].swizzle[chan] != chan) return false; } } return true; } /* Recognize the y = x - ffract(x) patterns from lowered ffloor. * It only works for the simple case when no swizzling is involved. */ static bool check_for_lowered_ffloor(nir_alu_instr *fadd) { if (!instr_has_only_trivial_swizzles(fadd)) return false; nir_alu_instr *fneg = NULL; nir_src x; for (unsigned i = 0; i < 2; i++) { nir_alu_instr *fadd_src_alu = nir_src_as_alu_instr(fadd->src[i].src); if (fadd_src_alu && fadd_src_alu->op == nir_op_fneg) { fneg = fadd_src_alu; x = fadd->src[1 - i].src; } } if (!fneg || !instr_has_only_trivial_swizzles(fneg)) return false; nir_alu_instr *ffract = nir_src_as_alu_instr(fneg->src[0].src); if (ffract && ffract->op == nir_op_ffract && nir_srcs_equal(ffract->src[0].src, x) && instr_has_only_trivial_swizzles(ffract)) return true; return false; } static bool lower_alu_instr(nir_builder *b, nir_alu_instr *alu) { const nir_op_info *info = &nir_op_infos[alu->op]; bool is_bool_only = alu->def.bit_size == 1; for (unsigned i = 0; i < info->num_inputs; i++) { if (alu->src[i].src.ssa->bit_size != 1) is_bool_only = false; } if (is_bool_only) { /* avoid lowering integers ops are used for booleans (ieq,ine,etc) */ return false; } b->cursor = nir_before_instr(&alu->instr); /* Replacement SSA value */ nir_def *rep = NULL; switch (alu->op) { case nir_op_mov: case nir_op_vec2: case nir_op_vec3: case nir_op_vec4: case nir_op_bcsel: /* These we expect to have integers but the opcode doesn't change */ break; case nir_op_b2i32: alu->op = nir_op_b2f32; break; case nir_op_i2f32: alu->op = nir_op_mov; break; case nir_op_u2f32: alu->op = nir_op_mov; break; case nir_op_f2i32: { alu->op = nir_op_ftrunc; /* If the source was already integer, then we did't need to truncate and * can switch it to a mov that can be copy-propagated away. */ nir_alu_instr *src_alu = nir_src_as_alu_instr(alu->src[0].src); if (src_alu) { switch (src_alu->op) { /* Check for the y = x - ffract(x) patterns from lowered ffloor. */ case nir_op_fadd: if (check_for_lowered_ffloor(src_alu)) alu->op = nir_op_mov; break; case nir_op_fround_even: case nir_op_fceil: case nir_op_ftrunc: case nir_op_ffloor: alu->op = nir_op_mov; break; default: break; } } break; } case nir_op_f2u32: alu->op = nir_op_ffloor; break; case nir_op_ilt: alu->op = nir_op_flt; break; case nir_op_ige: alu->op = nir_op_fge; break; case nir_op_ieq: alu->op = nir_op_feq; break; case nir_op_ine: alu->op = nir_op_fneu; break; case nir_op_ult: alu->op = nir_op_flt; break; case nir_op_uge: alu->op = nir_op_fge; break; case nir_op_iadd: alu->op = nir_op_fadd; break; case nir_op_isub: alu->op = nir_op_fsub; break; case nir_op_imul: alu->op = nir_op_fmul; break; case nir_op_idiv: { nir_def *x = nir_ssa_for_alu_src(b, alu, 0); nir_def *y = nir_ssa_for_alu_src(b, alu, 1); /* Hand-lower fdiv, since lower_int_to_float is after nir_opt_algebraic. */ if (b->shader->options->lower_fdiv) { rep = nir_ftrunc(b, nir_fmul(b, x, nir_frcp(b, y))); } else { rep = nir_ftrunc(b, nir_fdiv(b, x, y)); } break; } case nir_op_iabs: alu->op = nir_op_fabs; break; case nir_op_ineg: alu->op = nir_op_fneg; break; case nir_op_imax: alu->op = nir_op_fmax; break; case nir_op_imin: alu->op = nir_op_fmin; break; case nir_op_umax: alu->op = nir_op_fmax; break; case nir_op_umin: alu->op = nir_op_fmin; break; case nir_op_ball_iequal2: alu->op = nir_op_ball_fequal2; break; case nir_op_ball_iequal3: alu->op = nir_op_ball_fequal3; break; case nir_op_ball_iequal4: alu->op = nir_op_ball_fequal4; break; case nir_op_bany_inequal2: alu->op = nir_op_bany_fnequal2; break; case nir_op_bany_inequal3: alu->op = nir_op_bany_fnequal3; break; case nir_op_bany_inequal4: alu->op = nir_op_bany_fnequal4; break; case nir_op_i32csel_gt: alu->op = nir_op_fcsel_gt; break; case nir_op_i32csel_ge: alu->op = nir_op_fcsel_ge; break; default: assert(nir_alu_type_get_base_type(info->output_type) != nir_type_int && nir_alu_type_get_base_type(info->output_type) != nir_type_uint); for (unsigned i = 0; i < info->num_inputs; i++) { assert(nir_alu_type_get_base_type(info->input_types[i]) != nir_type_int && nir_alu_type_get_base_type(info->input_types[i]) != nir_type_uint); } return false; } if (rep) { /* We've emitted a replacement instruction */ nir_def_replace(&alu->def, rep); } return true; } static bool nir_lower_int_to_float_impl(nir_function_impl *impl) { bool progress = false; BITSET_WORD *float_types = NULL, *int_types = NULL; nir_builder b = nir_builder_create(impl); nir_index_ssa_defs(impl); float_types = calloc(BITSET_WORDS(impl->ssa_alloc), sizeof(BITSET_WORD)); int_types = calloc(BITSET_WORDS(impl->ssa_alloc), sizeof(BITSET_WORD)); nir_gather_types(impl, float_types, int_types); 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 && BITSET_TEST(int_types, load->def.index)) { for (unsigned i = 0; i < load->def.num_components; i++) load->value[i].f32 = load->value[i].i32; } break; } case nir_instr_type_intrinsic: case nir_instr_type_undef: case nir_instr_type_phi: case nir_instr_type_tex: break; default: nir_foreach_def(instr, assert_ssa_def_is_not_int, (void *)int_types); break; } } } if (progress) { nir_metadata_preserve(impl, nir_metadata_control_flow); } else { nir_metadata_preserve(impl, nir_metadata_all); } free(float_types); free(int_types); return progress; } bool nir_lower_int_to_float(nir_shader *shader) { bool progress = false; nir_foreach_function_impl(impl, shader) { if (nir_lower_int_to_float_impl(impl)) progress = true; } return progress; }