src/freedreno/ir3/ir3_cf.c

   1 /*
   2  * Copyright (C) 2019 Google.
   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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "ir3.h"
  25
  26 static bool
  27 is_fp16_conv(struct ir3_instruction *instr)
  28 {
  29         if (instr->opc != OPC_MOV)
  30                 return false;
  31
  32         struct ir3_register *dst = instr->regs[0];
  33         struct ir3_register *src = instr->regs[1];
  34
  35         /* disallow conversions that cannot be folded into
  36          * alu instructions:
  37          */
  38         if (dst->flags & (IR3_REG_EVEN | IR3_REG_POS_INF))
  39                 return false;
  40
  41         if (dst->flags & (IR3_REG_RELATIV | IR3_REG_ARRAY))
  42                 return false;
  43         if (src->flags & (IR3_REG_RELATIV | IR3_REG_ARRAY))
  44                 return false;
  45
  46         if (instr->cat1.src_type == TYPE_F32 &&
  47                         instr->cat1.dst_type == TYPE_F16)
  48                 return true;
  49
  50         if (instr->cat1.src_type == TYPE_F16 &&
  51                         instr->cat1.dst_type == TYPE_F32)
  52                 return true;
  53
  54         return false;
  55 }
  56
  57 static bool
  58 all_uses_fp16_conv(struct ir3 *ir, struct ir3_instruction *conv_src)
  59 {
  60         foreach_block (block, &ir->block_list) {
  61                 foreach_instr (instr, &block->instr_list) {
  62                         struct ir3_instruction *src;
  63                         foreach_ssa_src (src, instr) {
  64                                 if (src == conv_src && !is_fp16_conv(instr))
  65                                         return false;
  66                         }
  67                 }
  68         }
  69
  70         return true;
  71 }
  72
  73 static void
  74 rewrite_uses(struct ir3 *ir, struct ir3_instruction *conv,
  75                          struct ir3_instruction *replace)
  76 {
  77         foreach_block (block, &ir->block_list) {
  78                 foreach_instr (instr, &block->instr_list) {
  79                         struct ir3_instruction *src;
  80                         foreach_ssa_src_n (src, n, instr) {
  81                                 if (src == conv)
  82                                         instr->regs[n]->instr = replace;
  83                         }
  84                 }
  85         }
  86 }
  87
  88 static void
  89 try_conversion_folding(struct ir3 *ir, struct ir3_instruction *conv)
  90 {
  91         struct ir3_instruction *src;
  92
  93         if (!is_fp16_conv(conv))
  94                 return;
  95
  96         src = ssa(conv->regs[1]);
  97         if (!is_alu(src))
  98                 return;
  99
 100         /* avoid folding f2f32(f2f16) together, in cases where this is legal to
 101          * do (glsl) nir should have handled that for us already:
 102          */
 103         if (is_fp16_conv(src))
 104                 return;
 105
 106         switch (src->opc) {
 107         case OPC_SEL_B32:
 108         case OPC_SEL_B16:
 109         case OPC_MAX_F:
 110         case OPC_MIN_F:
 111         case OPC_SIGN_F:
 112         case OPC_ABSNEG_F:
 113                 return;
 114         default:
 115                 break;
 116         }
 117
 118         if (!all_uses_fp16_conv(ir, src))
 119                 return;
 120
 121         if (src->opc == OPC_MOV) {
 122                 if (src->cat1.dst_type == src->cat1.src_type) {
 123                         /* If we're folding a conversion into a bitwise move, we need to
 124                          * change the dst type to F32 to get the right behavior, since we
 125                          * could be moving a float with a u32.u32 move.
 126                          */
 127                         src->cat1.dst_type = conv->cat1.dst_type;
 128                         src->cat1.src_type = conv->cat1.src_type;
 129                 } else {
 130                         /* Otherwise, for typechanging movs, we can just change the dst
 131                          * type to F16 to collaps the two conversions.  For example
 132                          * cov.s32f32 follwed by cov.f32f16 becomes cov.s32f16.
 133                          */
 134                         src->cat1.dst_type = conv->cat1.dst_type;
 135                 }
 136         }
 137
 138         if (conv->regs[0]->flags & IR3_REG_HALF) {
 139                 src->regs[0]->flags |= IR3_REG_HALF;
 140         } else {
 141                 src->regs[0]->flags &= ~IR3_REG_HALF;
 142         }
 143
 144         rewrite_uses(ir, conv, src);
 145 }
 146
 147 void
 148 ir3_cf(struct ir3 *ir)
 149 {
 150         foreach_block_safe (block, &ir->block_list) {
 151                 foreach_instr_safe (instr, &block->instr_list) {
 152                         try_conversion_folding(ir, instr);
 153                 }
 154         }
 155 }