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