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
  33 static void
  34 lower_instr(nir_builder *bld, nir_alu_instr *alu, unsigned bit_size)
  35 {
  36    const nir_op op = alu->op;
  37    unsigned dst_bit_size = alu->dest.dest.ssa.bit_size;
  38
  39    bld->cursor = nir_before_instr(&alu->instr);
  40
  41    /* Convert each source to the requested bit-size */
  42    nir_ssa_def *srcs[NIR_MAX_VEC_COMPONENTS] = { NULL };
  43    for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++) {
  44       nir_ssa_def *src = nir_ssa_for_alu_src(bld, alu, i);
  45
  46       nir_alu_type type = nir_op_infos[op].input_types[i];
  47       if (nir_alu_type_get_type_size(type) == 0)
  48          src = nir_convert_to_bit_size(bld, src, type, bit_size);
  49
  50       if (i == 1 && (op == nir_op_ishl || op == nir_op_ishr || op == nir_op_ushr)) {
  51          assert(util_is_power_of_two_nonzero(dst_bit_size));
  52          src = nir_iand(bld, src, nir_imm_int(bld, dst_bit_size - 1));
  53       }
  54
  55       srcs[i] = src;
  56    }
  57
  58    /* Emit the lowered ALU instruction */
  59    nir_ssa_def *lowered_dst = NULL;
  60    if (op == nir_op_imul_high || op == nir_op_umul_high) {
  61       assert(dst_bit_size * 2 <= bit_size);
  62       nir_ssa_def *lowered_dst = nir_imul(bld, srcs[0], srcs[1]);
  63       if (nir_op_infos[op].output_type & nir_type_uint)
  64          lowered_dst = nir_ushr_imm(bld, lowered_dst, dst_bit_size);
  65       else
  66          lowered_dst = nir_ishr_imm(bld, lowered_dst, dst_bit_size);
  67    } else {
  68       lowered_dst = nir_build_alu_src_arr(bld, op, srcs);
  69    }
  70
  71
  72    /* Convert result back to the original bit-size */
  73    nir_alu_type type = nir_op_infos[op].output_type;
  74    nir_ssa_def *dst = nir_convert_to_bit_size(bld, lowered_dst, type, dst_bit_size);
  75    nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(dst));
  76 }
  77
  78 static bool
  79 lower_impl(nir_function_impl *impl,
  80            nir_lower_bit_size_callback callback,
  81            void *callback_data)
  82 {
  83    nir_builder b;
  84    nir_builder_init(&b, impl);
  85    bool progress = false;
  86
  87    nir_foreach_block(block, impl) {
  88       nir_foreach_instr_safe(instr, block) {
  89          if (instr->type != nir_instr_type_alu)
  90             continue;
  91
  92          nir_alu_instr *alu = nir_instr_as_alu(instr);
  93          assert(alu->dest.dest.is_ssa);
  94
  95          unsigned lower_bit_size = callback(alu, callback_data);
  96          if (lower_bit_size == 0)
  97             continue;
  98
  99          assert(lower_bit_size != alu->dest.dest.ssa.bit_size);
 100
 101          lower_instr(&b, alu, lower_bit_size);
 102          progress = true;
 103       }
 104    }
 105
 106    if (progress) {
 107       nir_metadata_preserve(impl, nir_metadata_block_index |
 108                                   nir_metadata_dominance);
 109    } else {
 110       nir_metadata_preserve(impl, nir_metadata_all);
 111    }
 112
 113    return progress;
 114 }
 115
 116 bool
 117 nir_lower_bit_size(nir_shader *shader,
 118                    nir_lower_bit_size_callback callback,
 119                    void *callback_data)
 120 {
 121    bool progress = false;
 122
 123    nir_foreach_function(function, shader) {
 124       if (function->impl)
 125          progress |= lower_impl(function->impl, callback, callback_data);
 126    }
 127
 128    return progress;
 129 }
 130
 131 static void
 132 split_phi(nir_builder *b, nir_phi_instr *phi)
 133 {
 134    nir_phi_instr *lowered[2] = {
 135       nir_phi_instr_create(b->shader),
 136       nir_phi_instr_create(b->shader)
 137    };
 138    int num_components = phi->dest.ssa.num_components;
 139    assert(phi->dest.ssa.bit_size == 64);
 140
 141    nir_foreach_phi_src(src, phi) {
 142       assert(num_components == src->src.ssa->num_components);
 143
 144       b->cursor = nir_before_src(&src->src, false);
 145
 146       nir_ssa_def *x = nir_unpack_64_2x32_split_x(b, src->src.ssa);
 147       nir_ssa_def *y = nir_unpack_64_2x32_split_y(b, src->src.ssa);
 148
 149       nir_phi_src *xsrc = rzalloc(lowered[0], nir_phi_src);
 150       xsrc->pred = src->pred;
 151       xsrc->src = nir_src_for_ssa(x);
 152       exec_list_push_tail(&lowered[0]->srcs, &xsrc->node);
 153
 154       nir_phi_src *ysrc = rzalloc(lowered[1], nir_phi_src);
 155       ysrc->pred = src->pred;
 156       ysrc->src = nir_src_for_ssa(y);
 157       exec_list_push_tail(&lowered[1]->srcs, &ysrc->node);
 158    }
 159
 160    nir_ssa_dest_init(&lowered[0]->instr, &lowered[0]->dest,
 161                      num_components, 32, NULL);
 162    nir_ssa_dest_init(&lowered[1]->instr, &lowered[1]->dest,
 163                      num_components, 32, NULL);
 164
 165    b->cursor = nir_before_instr(&phi->instr);
 166    nir_builder_instr_insert(b, &lowered[0]->instr);
 167    nir_builder_instr_insert(b, &lowered[1]->instr);
 168
 169    b->cursor = nir_after_phis(nir_cursor_current_block(b->cursor));
 170    nir_ssa_def *merged = nir_pack_64_2x32_split(b, &lowered[0]->dest.ssa, &lowered[1]->dest.ssa);
 171    nir_ssa_def_rewrite_uses(&phi->dest.ssa, nir_src_for_ssa(merged));
 172    nir_instr_remove(&phi->instr);
 173 }
 174
 175 static bool
 176 lower_64bit_phi_impl(nir_function_impl *impl)
 177 {
 178    nir_builder b;
 179    nir_builder_init(&b, impl);
 180    bool progress = false;
 181
 182    nir_foreach_block(block, impl) {
 183       nir_foreach_instr_safe(instr, block) {
 184          if (instr->type != nir_instr_type_phi)
 185             break;
 186
 187          nir_phi_instr *phi = nir_instr_as_phi(instr);
 188          assert(phi->dest.is_ssa);
 189
 190          if (phi->dest.ssa.bit_size <= 32)
 191             continue;
 192
 193          split_phi(&b, phi);
 194          progress = true;
 195       }
 196    }
 197
 198    if (progress) {
 199       nir_metadata_preserve(impl, nir_metadata_block_index |
 200                                   nir_metadata_dominance);
 201    } else {
 202       nir_metadata_preserve(impl, nir_metadata_all);
 203    }
 204
 205    return progress;
 206 }
 207
 208 bool
 209 nir_lower_64bit_phis(nir_shader *shader)
 210 {
 211    bool progress = false;
 212
 213    nir_foreach_function(function, shader) {
 214       if (function->impl)
 215          progress |= lower_64bit_phi_impl(function->impl);
 216    }
 217
 218    return progress;
 219 }