src/gallium/drivers/i915/i915_fpc_optimize.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2011 The Chromium OS authors.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL GOOGLE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "i915_reg.h"
  29 #include "i915_context.h"
  30 #include "i915_fpc.h"
  31
  32 #include "pipe/p_shader_tokens.h"
  33 #include "util/u_math.h"
  34 #include "util/u_memory.h"
  35 #include "util/u_string.h"
  36 #include "tgsi/tgsi_parse.h"
  37 #include "tgsi/tgsi_dump.h"
  38
  39 static boolean same_src_dst_reg(struct i915_full_src_register* s1, struct i915_full_dst_register* d1)
  40 {
  41    return (s1->Register.File == d1->Register.File &&
  42            s1->Register.Indirect == d1->Register.Indirect &&
  43            s1->Register.Dimension == d1->Register.Dimension &&
  44            s1->Register.Index == d1->Register.Index);
  45 }
  46
  47 static boolean same_dst_reg(struct i915_full_dst_register* d1, struct i915_full_dst_register* d2)
  48 {
  49    return (d1->Register.File == d2->Register.File &&
  50            d1->Register.Indirect == d2->Register.Indirect &&
  51            d1->Register.Dimension == d2->Register.Dimension &&
  52            d1->Register.Index == d2->Register.Index);
  53 }
  54
  55 static boolean same_src_reg(struct i915_full_src_register* d1, struct i915_full_src_register* d2)
  56 {
  57    return (d1->Register.File == d2->Register.File &&
  58            d1->Register.Indirect == d2->Register.Indirect &&
  59            d1->Register.Dimension == d2->Register.Dimension &&
  60            d1->Register.Index == d2->Register.Index &&
  61            d1->Register.Absolute == d2->Register.Absolute &&
  62            d1->Register.Negate == d2->Register.Negate);
  63 }
  64
  65 static boolean has_destination(unsigned opcode)
  66 {
  67    return (opcode != TGSI_OPCODE_NOP &&
  68            opcode != TGSI_OPCODE_KIL &&
  69            opcode != TGSI_OPCODE_RET);
  70 }
  71
  72 static boolean is_unswizzled(struct i915_full_src_register* r,
  73                              unsigned write_mask)
  74 {
  75    if ( write_mask & TGSI_WRITEMASK_X && r->Register.SwizzleX != TGSI_SWIZZLE_X)
  76       return FALSE;
  77    if ( write_mask & TGSI_WRITEMASK_Y && r->Register.SwizzleY != TGSI_SWIZZLE_Y)
  78       return FALSE;
  79    if ( write_mask & TGSI_WRITEMASK_Z && r->Register.SwizzleZ != TGSI_SWIZZLE_Z)
  80       return FALSE;
  81    if ( write_mask & TGSI_WRITEMASK_W && r->Register.SwizzleW != TGSI_SWIZZLE_W)
  82       return FALSE;
  83    return TRUE;
  84 }
  85
  86 static boolean op_commutes(unsigned opcode)
  87 {
  88    if (opcode == TGSI_OPCODE_ADD) return TRUE;
  89    if (opcode == TGSI_OPCODE_MUL) return TRUE;
  90    return FALSE;
  91 }
  92
  93 static unsigned op_neutral_element(unsigned opcode)
  94 {
  95    if (opcode == TGSI_OPCODE_ADD)
  96       return TGSI_SWIZZLE_ZERO;
  97    if (opcode == TGSI_OPCODE_MUL)
  98       return TGSI_SWIZZLE_ONE;
  99
 100    debug_printf("Unknown opcode %d\n",opcode);
 101    return TGSI_SWIZZLE_ZERO;
 102 }
 103
 104 /*
 105  * Sets the swizzle to the neutral element for the operation for the bits
 106  * of writemask which are set, swizzle to identity otherwise.
 107  */
 108 static void set_neutral_element_swizzle(struct i915_full_src_register* r,
 109                                         unsigned write_mask,
 110                                         unsigned neutral)
 111 {
 112    if ( write_mask & TGSI_WRITEMASK_X )
 113       r->Register.SwizzleX = neutral;
 114    else
 115       r->Register.SwizzleX = TGSI_SWIZZLE_X;
 116
 117    if ( write_mask & TGSI_WRITEMASK_Y )
 118       r->Register.SwizzleY = neutral;
 119    else
 120       r->Register.SwizzleY = TGSI_SWIZZLE_Y;
 121
 122    if ( write_mask & TGSI_WRITEMASK_Z )
 123       r->Register.SwizzleZ = neutral;
 124    else
 125       r->Register.SwizzleZ = TGSI_SWIZZLE_Z;
 126
 127    if ( write_mask & TGSI_WRITEMASK_W )
 128       r->Register.SwizzleW = neutral;
 129    else
 130       r->Register.SwizzleW = TGSI_SWIZZLE_W;
 131 }
 132
 133 /*
 134  * Optimize away things like:
 135  *    MUL OUT[0].xyz, TEMP[1], TEMP[2]
 136  *    MOV OUT[0].w, TEMP[2]
 137  * into:
 138  *    MUL OUT[0].xyzw, TEMP[1].xyz1, TEMP[2]
 139  * This is useful for optimizing texenv.
 140  */
 141 static void i915_fpc_optimize_mov_after_alu(union i915_full_token* current, union i915_full_token* next)
 142 {
 143    if ( current->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 144         next->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 145         op_commutes(current->FullInstruction.Instruction.Opcode) &&
 146         current->FullInstruction.Instruction.Saturate == next->FullInstruction.Instruction.Saturate &&
 147         next->FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 148         same_dst_reg(&next->FullInstruction.Dst[0], &current->FullInstruction.Dst[0]) &&
 149         same_src_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Src[1]) &&
 150         is_unswizzled(&current->FullInstruction.Src[0], current->FullInstruction.Dst[0].Register.WriteMask) &&
 151         is_unswizzled(&current->FullInstruction.Src[1], current->FullInstruction.Dst[0].Register.WriteMask) &&
 152         is_unswizzled(&next->FullInstruction.Src[0], next->FullInstruction.Dst[0].Register.WriteMask) )
 153    {
 154       next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP;
 155
 156       set_neutral_element_swizzle(&current->FullInstruction.Src[1], 0, 0);
 157       set_neutral_element_swizzle(&current->FullInstruction.Src[0],
 158                                   next->FullInstruction.Dst[0].Register.WriteMask,
 159                                   op_neutral_element(current->FullInstruction.Instruction.Opcode));
 160
 161       current->FullInstruction.Dst[0].Register.WriteMask = current->FullInstruction.Dst[0].Register.WriteMask |
 162                                                            next->FullInstruction.Dst[0].Register.WriteMask;
 163       return;
 164    }
 165
 166    if ( current->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 167         next->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 168         op_commutes(current->FullInstruction.Instruction.Opcode) &&
 169         current->FullInstruction.Instruction.Saturate == next->FullInstruction.Instruction.Saturate &&
 170         next->FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 171         same_dst_reg(&next->FullInstruction.Dst[0], &current->FullInstruction.Dst[0]) &&
 172         same_src_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Src[0]) &&
 173         is_unswizzled(&current->FullInstruction.Src[0], current->FullInstruction.Dst[0].Register.WriteMask) &&
 174         is_unswizzled(&current->FullInstruction.Src[1], current->FullInstruction.Dst[0].Register.WriteMask) &&
 175         is_unswizzled(&next->FullInstruction.Src[0], next->FullInstruction.Dst[0].Register.WriteMask) )
 176    {
 177       next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP;
 178
 179       set_neutral_element_swizzle(&current->FullInstruction.Src[0], 0, 0);
 180       set_neutral_element_swizzle(&current->FullInstruction.Src[1],
 181                                   next->FullInstruction.Dst[0].Register.WriteMask,
 182                                   op_neutral_element(current->FullInstruction.Instruction.Opcode));
 183
 184       current->FullInstruction.Dst[0].Register.WriteMask = current->FullInstruction.Dst[0].Register.WriteMask |
 185                                                            next->FullInstruction.Dst[0].Register.WriteMask;
 186       return;
 187    }
 188 }
 189
 190 /*
 191  * Optimize away things like:
 192  *    *** TEMP[0], TEMP[1], TEMP[2]
 193  *    MOV OUT[0] TEMP[0]
 194  * into:
 195  *    *** OUT[0], TEMP[1], TEMP[2]
 196  */
 197 static void i915_fpc_optimize_useless_mov(union i915_full_token* current, union i915_full_token* next)
 198 {
 199    if ( current->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 200         next->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 201         next->FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 202         has_destination(current->FullInstruction.Instruction.Opcode) &&
 203         next->FullInstruction.Instruction.Saturate == TGSI_SAT_NONE &&
 204         next->FullInstruction.Src[0].Register.Absolute == 0 &&
 205         next->FullInstruction.Src[0].Register.Negate == 0 &&
 206         is_unswizzled(&next->FullInstruction.Src[0], next->FullInstruction.Dst[0].Register.WriteMask) &&
 207         current->FullInstruction.Dst[0].Register.WriteMask == next->FullInstruction.Dst[0].Register.WriteMask &&
 208         same_src_dst_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Dst[0]) )
 209    {
 210       next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP;
 211
 212       current->FullInstruction.Dst[0] = next->FullInstruction.Dst[0];
 213       return;
 214    }
 215 }
 216
 217 static void copy_src_reg(struct i915_src_register* o, const struct tgsi_src_register* i)
 218 {
 219    o->File      = i->File;
 220    o->Indirect  = i->Indirect;
 221    o->Dimension = i->Dimension;
 222    o->Index     = i->Index;
 223    o->SwizzleX  = i->SwizzleX;
 224    o->SwizzleY  = i->SwizzleY;
 225    o->SwizzleZ  = i->SwizzleZ;
 226    o->SwizzleW  = i->SwizzleW;
 227    o->Absolute  = i->Absolute;
 228    o->Negate    = i->Negate;
 229 }
 230
 231 static void copy_dst_reg(struct i915_dst_register* o, const struct tgsi_dst_register* i)
 232 {
 233    o->File      = i->File;
 234    o->WriteMask = i->WriteMask;
 235    o->Indirect  = i->Indirect;
 236    o->Dimension = i->Dimension;
 237    o->Index     = i->Index;
 238 }
 239
 240 static void copy_instruction(struct i915_full_instruction* o, const struct tgsi_full_instruction* i)
 241 {
 242    memcpy(&o->Instruction, &i->Instruction, sizeof(o->Instruction));
 243    memcpy(&o->Texture, &i->Texture, sizeof(o->Texture));
 244
 245    copy_dst_reg(&o->Dst[0].Register, &i->Dst[0].Register);
 246
 247    copy_src_reg(&o->Src[0].Register, &i->Src[0].Register);
 248    copy_src_reg(&o->Src[1].Register, &i->Src[1].Register);
 249    copy_src_reg(&o->Src[2].Register, &i->Src[2].Register);
 250 }
 251
 252 static void copy_token(union i915_full_token* o, union tgsi_full_token* i)
 253 {
 254    if (i->Token.Type != TGSI_TOKEN_TYPE_INSTRUCTION)
 255       memcpy(o, i, sizeof(*o));
 256    else
 257       copy_instruction(&o->FullInstruction, &i->FullInstruction);
 258
 259 }
 260
 261 struct i915_token_list* i915_optimize(const struct tgsi_token *tokens)
 262 {
 263    struct i915_token_list *out_tokens = MALLOC(sizeof(struct i915_token_list));
 264    struct tgsi_parse_context parse;
 265    int i = 0;
 266
 267    out_tokens->NumTokens = 0;
 268
 269    /* Count the tokens */
 270    tgsi_parse_init( &parse, tokens );
 271    while( !tgsi_parse_end_of_tokens( &parse ) ) {
 272       tgsi_parse_token( &parse );
 273       out_tokens->NumTokens++;
 274    }
 275    tgsi_parse_free (&parse);
 276
 277    /* Allocate our tokens */
 278    out_tokens->Tokens = MALLOC(sizeof(union i915_full_token) * out_tokens->NumTokens);
 279
 280    tgsi_parse_init( &parse, tokens );
 281    while( !tgsi_parse_end_of_tokens( &parse ) ) {
 282       tgsi_parse_token( &parse );
 283       copy_token(&out_tokens->Tokens[i] , &parse.FullToken);
 284
 285       if (i > 0) {
 286          i915_fpc_optimize_useless_mov(&out_tokens->Tokens[i-1], &out_tokens->Tokens[i]);
 287          i915_fpc_optimize_mov_after_alu(&out_tokens->Tokens[i-1], &out_tokens->Tokens[i]);
 288       }
 289       i++;
 290    }
 291    tgsi_parse_free (&parse);
 292
 293    return out_tokens;
 294 }
 295
 296 void i915_optimize_free(struct i915_token_list* tokens)
 297 {
 298    free(tokens->Tokens);
 299    free(tokens);
 300 }
 301
 302