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_KILL_IF &&
  69            opcode != TGSI_OPCODE_KILL &&
  70            opcode != TGSI_OPCODE_END &&
  71            opcode != TGSI_OPCODE_RET);
  72 }
  73
  74 static boolean is_unswizzled(struct i915_full_src_register* r,
  75                              unsigned write_mask)
  76 {
  77    if ( write_mask & TGSI_WRITEMASK_X && r->Register.SwizzleX != TGSI_SWIZZLE_X)
  78       return FALSE;
  79    if ( write_mask & TGSI_WRITEMASK_Y && r->Register.SwizzleY != TGSI_SWIZZLE_Y)
  80       return FALSE;
  81    if ( write_mask & TGSI_WRITEMASK_Z && r->Register.SwizzleZ != TGSI_SWIZZLE_Z)
  82       return FALSE;
  83    if ( write_mask & TGSI_WRITEMASK_W && r->Register.SwizzleW != TGSI_SWIZZLE_W)
  84       return FALSE;
  85    return TRUE;
  86 }
  87
  88 static boolean op_commutes(unsigned opcode)
  89 {
  90    switch(opcode)
  91    {
  92       case TGSI_OPCODE_ADD:
  93       case TGSI_OPCODE_MUL:
  94       case TGSI_OPCODE_DP2:
  95       case TGSI_OPCODE_DP3:
  96       case TGSI_OPCODE_DP4:
  97          return TRUE;
  98    }
  99    return FALSE;
 100 }
 101
 102 static unsigned op_neutral_element(unsigned opcode)
 103 {
 104    switch(opcode)
 105    {
 106       case TGSI_OPCODE_ADD:
 107          return TGSI_SWIZZLE_ZERO;
 108       case TGSI_OPCODE_MUL:
 109       case TGSI_OPCODE_DP2:
 110       case TGSI_OPCODE_DP3:
 111       case TGSI_OPCODE_DP4:
 112          return TGSI_SWIZZLE_ONE;
 113    }
 114
 115    debug_printf("Unknown opcode %d\n",opcode);
 116    return TGSI_SWIZZLE_ZERO;
 117 }
 118
 119 /*
 120  * Sets the swizzle to the neutral element for the operation for the bits
 121  * of writemask which are set, swizzle to identity otherwise.
 122  */
 123 static void set_neutral_element_swizzle(struct i915_full_src_register* r,
 124                                         unsigned write_mask,
 125                                         unsigned neutral)
 126 {
 127    if ( write_mask & TGSI_WRITEMASK_X )
 128       r->Register.SwizzleX = neutral;
 129    else
 130       r->Register.SwizzleX = TGSI_SWIZZLE_X;
 131
 132    if ( write_mask & TGSI_WRITEMASK_Y )
 133       r->Register.SwizzleY = neutral;
 134    else
 135       r->Register.SwizzleY = TGSI_SWIZZLE_Y;
 136
 137    if ( write_mask & TGSI_WRITEMASK_Z )
 138       r->Register.SwizzleZ = neutral;
 139    else
 140       r->Register.SwizzleZ = TGSI_SWIZZLE_Z;
 141
 142    if ( write_mask & TGSI_WRITEMASK_W )
 143       r->Register.SwizzleW = neutral;
 144    else
 145       r->Register.SwizzleW = TGSI_SWIZZLE_W;
 146 }
 147
 148 static void copy_src_reg(struct i915_src_register* o, const struct tgsi_src_register* i)
 149 {
 150    o->File      = i->File;
 151    o->Indirect  = i->Indirect;
 152    o->Dimension = i->Dimension;
 153    o->Index     = i->Index;
 154    o->SwizzleX  = i->SwizzleX;
 155    o->SwizzleY  = i->SwizzleY;
 156    o->SwizzleZ  = i->SwizzleZ;
 157    o->SwizzleW  = i->SwizzleW;
 158    o->Absolute  = i->Absolute;
 159    o->Negate    = i->Negate;
 160 }
 161
 162 static void copy_dst_reg(struct i915_dst_register* o, const struct tgsi_dst_register* i)
 163 {
 164    o->File      = i->File;
 165    o->WriteMask = i->WriteMask;
 166    o->Indirect  = i->Indirect;
 167    o->Dimension = i->Dimension;
 168    o->Index     = i->Index;
 169 }
 170
 171 static void copy_instruction(struct i915_full_instruction* o, const struct tgsi_full_instruction* i)
 172 {
 173    memcpy(&o->Instruction, &i->Instruction, sizeof(o->Instruction));
 174    memcpy(&o->Texture, &i->Texture, sizeof(o->Texture));
 175
 176    copy_dst_reg(&o->Dst[0].Register, &i->Dst[0].Register);
 177
 178    copy_src_reg(&o->Src[0].Register, &i->Src[0].Register);
 179    copy_src_reg(&o->Src[1].Register, &i->Src[1].Register);
 180    copy_src_reg(&o->Src[2].Register, &i->Src[2].Register);
 181 }
 182
 183 static void copy_token(union i915_full_token* o, union tgsi_full_token* i)
 184 {
 185    if (i->Token.Type != TGSI_TOKEN_TYPE_INSTRUCTION)
 186       memcpy(o, i, sizeof(*o));
 187    else
 188       copy_instruction(&o->FullInstruction, &i->FullInstruction);
 189
 190 }
 191
 192 /*
 193  * Optimize away things like:
 194  *    MUL OUT[0].xyz, TEMP[1], TEMP[2]
 195  *    MOV OUT[0].w, TEMP[2]
 196  * into:
 197  *    MUL OUT[0].xyzw, TEMP[1].xyz1, TEMP[2]
 198  * This is useful for optimizing texenv.
 199  */
 200 static void i915_fpc_optimize_mov_after_alu(union i915_full_token* current, union i915_full_token* next)
 201 {
 202    if ( current->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 203         next->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 204         op_commutes(current->FullInstruction.Instruction.Opcode) &&
 205         current->FullInstruction.Instruction.Saturate == next->FullInstruction.Instruction.Saturate &&
 206         next->FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 207         same_dst_reg(&next->FullInstruction.Dst[0], &current->FullInstruction.Dst[0]) &&
 208         same_src_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Src[1]) &&
 209         !same_src_dst_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Dst[0]) &&
 210         is_unswizzled(&current->FullInstruction.Src[0], current->FullInstruction.Dst[0].Register.WriteMask) &&
 211         is_unswizzled(&current->FullInstruction.Src[1], current->FullInstruction.Dst[0].Register.WriteMask) &&
 212         is_unswizzled(&next->FullInstruction.Src[0], next->FullInstruction.Dst[0].Register.WriteMask) )
 213    {
 214       next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP;
 215
 216       set_neutral_element_swizzle(&current->FullInstruction.Src[1], 0, 0);
 217       set_neutral_element_swizzle(&current->FullInstruction.Src[0],
 218                                   next->FullInstruction.Dst[0].Register.WriteMask,
 219                                   op_neutral_element(current->FullInstruction.Instruction.Opcode));
 220
 221       current->FullInstruction.Dst[0].Register.WriteMask = current->FullInstruction.Dst[0].Register.WriteMask |
 222                                                            next->FullInstruction.Dst[0].Register.WriteMask;
 223       return;
 224    }
 225
 226    if ( current->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 227         next->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 228         op_commutes(current->FullInstruction.Instruction.Opcode) &&
 229         current->FullInstruction.Instruction.Saturate == next->FullInstruction.Instruction.Saturate &&
 230         next->FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 231         same_dst_reg(&next->FullInstruction.Dst[0], &current->FullInstruction.Dst[0]) &&
 232         same_src_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Src[0]) &&
 233         !same_src_dst_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Dst[0]) &&
 234         is_unswizzled(&current->FullInstruction.Src[0], current->FullInstruction.Dst[0].Register.WriteMask) &&
 235         is_unswizzled(&current->FullInstruction.Src[1], current->FullInstruction.Dst[0].Register.WriteMask) &&
 236         is_unswizzled(&next->FullInstruction.Src[0], next->FullInstruction.Dst[0].Register.WriteMask) )
 237    {
 238       next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP;
 239
 240       set_neutral_element_swizzle(&current->FullInstruction.Src[0], 0, 0);
 241       set_neutral_element_swizzle(&current->FullInstruction.Src[1],
 242                                   next->FullInstruction.Dst[0].Register.WriteMask,
 243                                   op_neutral_element(current->FullInstruction.Instruction.Opcode));
 244
 245       current->FullInstruction.Dst[0].Register.WriteMask = current->FullInstruction.Dst[0].Register.WriteMask |
 246                                                            next->FullInstruction.Dst[0].Register.WriteMask;
 247       return;
 248    }
 249 }
 250
 251 /*
 252  * Optimize away things like:
 253  *    MOV TEMP[0].xyz TEMP[0].xyzx
 254  * into:
 255  *    NOP
 256  */
 257 static boolean i915_fpc_useless_mov(union tgsi_full_token* tgsi_current)
 258 {
 259    union i915_full_token current;
 260    copy_token(&current , tgsi_current);
 261    if ( current.Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 262         current.FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 263         has_destination(current.FullInstruction.Instruction.Opcode) &&
 264         current.FullInstruction.Instruction.Saturate == TGSI_SAT_NONE &&
 265         current.FullInstruction.Src[0].Register.Absolute == 0 &&
 266         current.FullInstruction.Src[0].Register.Negate == 0 &&
 267         is_unswizzled(&current.FullInstruction.Src[0], current.FullInstruction.Dst[0].Register.WriteMask) &&
 268         same_src_dst_reg(&current.FullInstruction.Src[0], &current.FullInstruction.Dst[0]) )
 269    {
 270       return TRUE;
 271    }
 272    return FALSE;
 273 }
 274
 275 /*
 276  * Optimize away things like:
 277  *    *** TEMP[0], TEMP[1], TEMP[2]
 278  *    MOV OUT[0] TEMP[0]
 279  * into:
 280  *    *** OUT[0], TEMP[1], TEMP[2]
 281  */
 282 static void i915_fpc_optimize_useless_mov_after_inst(union i915_full_token* current, union i915_full_token* next)
 283 {
 284    if ( current->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 285         next->Token.Type == TGSI_TOKEN_TYPE_INSTRUCTION  &&
 286         next->FullInstruction.Instruction.Opcode == TGSI_OPCODE_MOV &&
 287         has_destination(current->FullInstruction.Instruction.Opcode) &&
 288         next->FullInstruction.Instruction.Saturate == TGSI_SAT_NONE &&
 289         next->FullInstruction.Src[0].Register.Absolute == 0 &&
 290         next->FullInstruction.Src[0].Register.Negate == 0 &&
 291         next->FullInstruction.Dst[0].Register.File == TGSI_FILE_OUTPUT &&
 292         is_unswizzled(&next->FullInstruction.Src[0], next->FullInstruction.Dst[0].Register.WriteMask) &&
 293         current->FullInstruction.Dst[0].Register.WriteMask == next->FullInstruction.Dst[0].Register.WriteMask &&
 294         same_src_dst_reg(&next->FullInstruction.Src[0], &current->FullInstruction.Dst[0]) )
 295    {
 296       next->FullInstruction.Instruction.Opcode = TGSI_OPCODE_NOP;
 297
 298       current->FullInstruction.Dst[0] = next->FullInstruction.Dst[0];
 299       return;
 300    }
 301 }
 302
 303 struct i915_token_list* i915_optimize(const struct tgsi_token *tokens)
 304 {
 305    struct i915_token_list *out_tokens = MALLOC(sizeof(struct i915_token_list));
 306    struct tgsi_parse_context parse;
 307    int i = 0;
 308
 309    out_tokens->NumTokens = 0;
 310
 311    /* Count the tokens */
 312    tgsi_parse_init( &parse, tokens );
 313    while( !tgsi_parse_end_of_tokens( &parse ) ) {
 314       tgsi_parse_token( &parse );
 315       out_tokens->NumTokens++;
 316    }
 317    tgsi_parse_free (&parse);
 318
 319    /* Allocate our tokens */
 320    out_tokens->Tokens = MALLOC(sizeof(union i915_full_token) * out_tokens->NumTokens);
 321
 322    tgsi_parse_init( &parse, tokens );
 323    while( !tgsi_parse_end_of_tokens( &parse ) ) {
 324       tgsi_parse_token( &parse );
 325
 326       if (i915_fpc_useless_mov(&parse.FullToken)) {
 327          out_tokens->NumTokens--;
 328          continue;
 329       }
 330
 331       copy_token(&out_tokens->Tokens[i] , &parse.FullToken);
 332
 333       if (i > 0) {
 334          i915_fpc_optimize_useless_mov_after_inst(&out_tokens->Tokens[i-1], &out_tokens->Tokens[i]);
 335          i915_fpc_optimize_mov_after_alu(&out_tokens->Tokens[i-1], &out_tokens->Tokens[i]);
 336       }
 337       i++;
 338    }
 339    tgsi_parse_free (&parse);
 340
 341    return out_tokens;
 342 }
 343
 344 void i915_optimize_free(struct i915_token_list* tokens)
 345 {
 346    free(tokens->Tokens);
 347    free(tokens);
 348 }
 349
 350