src/compiler/nir/nir_lower_bit_size.c

   1 /*
   2  * Copyright © 2018 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "nir_builder.h"
  25
  26 /**
  27  * Some ALU operations may not be supported in hardware in specific bit-sizes.
  28  * This pass allows implementations to selectively lower such operations to
  29  * a bit-size that is supported natively and then converts the result back to
  30  * the original bit-size.
  31  */
  32 static nir_ssa_def *
  33 convert_to_bit_size(nir_builder *bld,
  34                     nir_ssa_def *src,
  35                     nir_alu_type type,
  36                     unsigned bit_size)
  37 {
  38    nir_alu_type base_type = nir_alu_type_get_base_type(type);
  39    nir_alu_type lowered_type = bit_size | base_type;
  40
  41    nir_op opcode =
  42       nir_type_conversion_op(type, lowered_type, nir_rounding_mode_undef);
  43
  44    return nir_build_alu(bld, opcode, src, NULL, NULL, NULL);
  45 }
  46
  47 static void
  48 lower_instr(nir_builder *bld, nir_alu_instr *alu, unsigned bit_size)
  49 {
  50    const nir_op op = alu->op;
  51    unsigned dst_bit_size = alu->dest.dest.ssa.bit_size;
  52
  53    bld->cursor = nir_before_instr(&alu->instr);
  54
  55    /* Convert each source to the requested bit-size */
  56    nir_ssa_def *srcs[4] = { NULL, NULL, NULL, NULL };
  57    for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++) {
  58       nir_ssa_def *src = nir_ssa_for_alu_src(bld, alu, i);
  59
  60       nir_alu_type type = nir_op_infos[op].input_types[i];
  61       if (nir_alu_type_get_type_size(type) == 0)
  62          src = convert_to_bit_size(bld, src, type, bit_size);
  63
  64       if (i == 1 && (op == nir_op_ishl || op == nir_op_ishr || op == nir_op_ushr)) {
  65          assert(util_is_power_of_two_nonzero(dst_bit_size));
  66          src = nir_iand(bld, src, nir_imm_int(bld, dst_bit_size - 1));
  67       }
  68
  69       srcs[i] = src;
  70    }
  71
  72    /* Emit the lowered ALU instruction */
  73    nir_ssa_def *lowered_dst = NULL;
  74    if (op == nir_op_imul_high || op == nir_op_umul_high) {
  75       assert(dst_bit_size * 2 <= bit_size);
  76       nir_ssa_def *lowered_dst = nir_imul(bld, srcs[0], srcs[1]);
  77       if (nir_op_infos[op].output_type & nir_type_uint)
  78          lowered_dst = nir_ushr(bld, lowered_dst, nir_imm_int(bld, dst_bit_size));
  79       else
  80          lowered_dst = nir_ishr(bld, lowered_dst, nir_imm_int(bld, dst_bit_size));
  81    } else {
  82       lowered_dst = nir_build_alu(bld, op, srcs[0], srcs[1], srcs[2], srcs[3]);
  83    }
  84
  85
  86    /* Convert result back to the original bit-size */
  87    nir_alu_type type = nir_op_infos[op].output_type;
  88    nir_ssa_def *dst = convert_to_bit_size(bld, lowered_dst, type, dst_bit_size);
  89    nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(dst));
  90 }
  91
  92 static bool
  93 lower_impl(nir_function_impl *impl,
  94            nir_lower_bit_size_callback callback,
  95            void *callback_data)
  96 {
  97    nir_builder b;
  98    nir_builder_init(&b, impl);
  99    bool progress = false;
 100
 101    nir_foreach_block(block, impl) {
 102       nir_foreach_instr_safe(instr, block) {
 103          if (instr->type != nir_instr_type_alu)
 104             continue;
 105
 106          nir_alu_instr *alu = nir_instr_as_alu(instr);
 107          assert(alu->dest.dest.is_ssa);
 108
 109          unsigned lower_bit_size = callback(alu, callback_data);
 110          if (lower_bit_size == 0)
 111             continue;
 112
 113          assert(lower_bit_size != alu->dest.dest.ssa.bit_size);
 114
 115          lower_instr(&b, alu, lower_bit_size);
 116          progress = true;
 117       }
 118    }
 119
 120    if (progress) {
 121       nir_metadata_preserve(impl, nir_metadata_block_index |
 122                                   nir_metadata_dominance);
 123    }
 124
 125    return progress;
 126 }
 127
 128 bool
 129 nir_lower_bit_size(nir_shader *shader,
 130                    nir_lower_bit_size_callback callback,
 131                    void *callback_data)
 132 {
 133    bool progress = false;
 134
 135    nir_foreach_function(function, shader) {
 136       if (function->impl)
 137          progress |= lower_impl(function->impl, callback, callback_data);
 138    }
 139
 140    return progress;
 141 }