src/gallium/drivers/vc4/vc4_opt_vpm.c

   1 /*
   2  * Copyright © 2014 Broadcom
   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 /**
  25  * @file vc4_opt_vpm.c
  26  *
  27  * This modifies instructions that:
  28  * 1. exclusively consume a value read from the VPM to directly read the VPM if
  29  *    other operands allow it.
  30  * 2. generate the value consumed by a VPM write to write directly into the VPM.
  31  */
  32
  33 #include "vc4_qir.h"
  34
  35 bool
  36 qir_opt_vpm(struct vc4_compile *c)
  37 {
  38         if (c->stage == QSTAGE_FRAG)
  39                 return false;
  40
  41         bool progress = false;
  42         struct qinst *vpm_writes[64] = { 0 };
  43         uint32_t use_count[c->num_temps];
  44         uint32_t vpm_write_count = 0;
  45         memset(&use_count, 0, sizeof(use_count));
  46
  47         list_for_each_entry(struct qinst, inst, &c->instructions, link) {
  48                 switch (inst->dst.file) {
  49                 case QFILE_VPM:
  50                         vpm_writes[vpm_write_count++] = inst;
  51                         break;
  52                 default:
  53                         break;
  54                 }
  55
  56                 for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
  57                         if (inst->src[i].file == QFILE_TEMP) {
  58                                 uint32_t temp = inst->src[i].index;
  59                                 use_count[temp]++;
  60                         }
  61                 }
  62         }
  63
  64         /* For instructions reading from a temporary that contains a VPM read
  65          * result, try to move the instruction up in place of the VPM read.
  66          */
  67         list_for_each_entry(struct qinst, inst, &c->instructions, link) {
  68                 if (!inst)
  69                         continue;
  70
  71                 if (qir_depends_on_flags(inst) || inst->sf)
  72                         continue;
  73
  74                 if (qir_has_side_effects(c, inst) ||
  75                     qir_has_side_effect_reads(c, inst) ||
  76                     qir_is_tex(inst))
  77                         continue;
  78
  79                 for (int j = 0; j < qir_get_op_nsrc(inst->op); j++) {
  80                         if (inst->src[j].file != QFILE_TEMP ||
  81                             inst->src[j].pack)
  82                                 continue;
  83
  84                         uint32_t temp = inst->src[j].index;
  85
  86                         /* Since VPM reads pull from a FIFO, we only get to
  87                          * read each VPM entry once (unless we reset the read
  88                          * pointer).  That means we can't copy-propagate a VPM
  89                          * read to multiple locations.
  90                          */
  91                         if (use_count[temp] != 1)
  92                                 continue;
  93
  94                         struct qinst *mov = c->defs[temp];
  95                         if (!mov ||
  96                             (mov->op != QOP_MOV &&
  97                              mov->op != QOP_FMOV &&
  98                              mov->op != QOP_MMOV) ||
  99                             mov->src[0].file != QFILE_VPM) {
 100                                 continue;
 101                         }
 102
 103                         uint32_t temps = 0;
 104                         for (int k = 0; k < qir_get_op_nsrc(inst->op); k++) {
 105                                 if (inst->src[k].file == QFILE_TEMP)
 106                                         temps++;
 107                         }
 108
 109                         /* The instruction is safe to reorder if its other
 110                          * sources are independent of previous instructions
 111                          */
 112                         if (temps == 1) {
 113                                 list_del(&inst->link);
 114                                 inst->src[j] = mov->src[0];
 115                                 list_replace(&mov->link, &inst->link);
 116                                 c->defs[temp] = NULL;
 117                                 free(mov);
 118                                 progress = true;
 119                                 break;
 120                         }
 121                 }
 122         }
 123
 124         for (int i = 0; i < vpm_write_count; i++) {
 125                 if (!qir_is_raw_mov(vpm_writes[i]) ||
 126                     vpm_writes[i]->src[0].file != QFILE_TEMP) {
 127                         continue;
 128                 }
 129
 130                 uint32_t temp = vpm_writes[i]->src[0].index;
 131                 if (use_count[temp] != 1)
 132                         continue;
 133
 134                 struct qinst *inst = c->defs[temp];
 135                 if (!inst)
 136                         continue;
 137
 138                 if (qir_depends_on_flags(inst) || inst->sf)
 139                         continue;
 140
 141                 if (qir_has_side_effects(c, inst) ||
 142                     qir_has_side_effect_reads(c, inst)) {
 143                         continue;
 144                 }
 145
 146                 /* Move the generating instruction to the end of the program
 147                  * to maintain the order of the VPM writes.
 148                  */
 149                 assert(!vpm_writes[i]->sf);
 150                 list_del(&inst->link);
 151                 list_addtail(&inst->link, &vpm_writes[i]->link);
 152                 qir_remove_instruction(c, vpm_writes[i]);
 153
 154                 c->defs[inst->dst.index] = NULL;
 155                 inst->dst.file = QFILE_VPM;
 156                 inst->dst.index = 0;
 157
 158                 progress = true;
 159         }
 160
 161         return progress;
 162 }