src/mesa/drivers/dri/i965/brw_eu.c

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a 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, sublicense, 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
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keith@tungstengraphics.com>
  30   */
  31
  32
  33 #include "brw_context.h"
  34 #include "brw_defines.h"
  35 #include "brw_eu.h"
  36
  37
  38
  39 /* How does predicate control work when execution_size != 8?  Do I
  40  * need to test/set for 0xffff when execution_size is 16?
  41  */
  42 void brw_set_predicate_control_flag_value( struct brw_compile *p, GLuint value )
  43 {
  44    p->current->header.predicate_control = BRW_PREDICATE_NONE;
  45
  46    if (value != 0xff) {
  47       if (value != p->flag_value) {
  48          brw_push_insn_state(p);
  49          brw_MOV(p, brw_flag_reg(), brw_imm_uw(value));
  50          p->flag_value = value;
  51          brw_pop_insn_state(p);
  52       }
  53
  54       p->current->header.predicate_control = BRW_PREDICATE_NORMAL;
  55    }
  56 }
  57
  58 void brw_set_predicate_control( struct brw_compile *p, GLuint pc )
  59 {
  60    p->current->header.predicate_control = pc;
  61 }
  62
  63 void brw_set_conditionalmod( struct brw_compile *p, GLuint conditional )
  64 {
  65    p->current->header.destreg__conditionalmod = conditional;
  66 }
  67
  68 void brw_set_access_mode( struct brw_compile *p, GLuint access_mode )
  69 {
  70    p->current->header.access_mode = access_mode;
  71 }
  72
  73 void brw_set_compression_control( struct brw_compile *p, GLboolean compression_control )
  74 {
  75    p->compressed = (compression_control == BRW_COMPRESSION_COMPRESSED);
  76
  77    if (p->brw->intel.gen >= 6) {
  78       /* Since we don't use the 32-wide support in gen6, we translate
  79        * the pre-gen6 compression control here.
  80        */
  81       switch (compression_control) {
  82       case BRW_COMPRESSION_NONE:
  83          /* This is the "use the first set of bits of dmask/vmask/arf
  84           * according to execsize" option.
  85           */
  86          p->current->header.compression_control = GEN6_COMPRESSION_1Q;
  87          break;
  88       case BRW_COMPRESSION_2NDHALF:
  89          /* For 8-wide, this is "use the second set of 8 bits." */
  90          p->current->header.compression_control = GEN6_COMPRESSION_2Q;
  91          break;
  92       case BRW_COMPRESSION_COMPRESSED:
  93          /* For 16-wide instruction compression, use the first set of 16 bits
  94           * since we don't do 32-wide dispatch.
  95           */
  96          p->current->header.compression_control = GEN6_COMPRESSION_1H;
  97          break;
  98       default:
  99          assert(!"not reached");
 100          p->current->header.compression_control = GEN6_COMPRESSION_1H;
 101          break;
 102       }
 103    } else {
 104       p->current->header.compression_control = compression_control;
 105    }
 106 }
 107
 108 void brw_set_mask_control( struct brw_compile *p, GLuint value )
 109 {
 110    p->current->header.mask_control = value;
 111 }
 112
 113 void brw_set_saturate( struct brw_compile *p, GLuint value )
 114 {
 115    p->current->header.saturate = value;
 116 }
 117
 118 void brw_set_acc_write_control(struct brw_compile *p, GLuint value)
 119 {
 120    if (p->brw->intel.gen >= 6)
 121       p->current->header.acc_wr_control = value;
 122 }
 123
 124 void brw_push_insn_state( struct brw_compile *p )
 125 {
 126    assert(p->current != &p->stack[BRW_EU_MAX_INSN_STACK-1]);
 127    memcpy(p->current+1, p->current, sizeof(struct brw_instruction));
 128    p->compressed_stack[p->current - p->stack] = p->compressed;
 129    p->current++;
 130 }
 131
 132 void brw_pop_insn_state( struct brw_compile *p )
 133 {
 134    assert(p->current != p->stack);
 135    p->current--;
 136    p->compressed = p->compressed_stack[p->current - p->stack];
 137 }
 138
 139
 140 /***********************************************************************
 141  */
 142 void brw_init_compile( struct brw_context *brw, struct brw_compile *p )
 143 {
 144    p->brw = brw;
 145    p->nr_insn = 0;
 146    p->current = p->stack;
 147    p->compressed = false;
 148    memset(p->current, 0, sizeof(p->current[0]));
 149
 150    /* Some defaults?
 151     */
 152    brw_set_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
 153    brw_set_saturate(p, 0);
 154    brw_set_compression_control(p, BRW_COMPRESSION_NONE);
 155    brw_set_predicate_control_flag_value(p, 0xff);
 156 }
 157
 158
 159 const GLuint *brw_get_program( struct brw_compile *p,
 160                                GLuint *sz )
 161 {
 162    GLuint i;
 163
 164    for (i = 0; i < 8; i++)
 165       brw_NOP(p);
 166
 167    *sz = p->nr_insn * sizeof(struct brw_instruction);
 168    return (const GLuint *)p->store;
 169 }
 170
 171
 172
 173 /**
 174  * Subroutine calls require special attention.
 175  * Mesa instructions may be expanded into multiple hardware instructions
 176  * so the prog_instruction::BranchTarget field can't be used as an index
 177  * into the hardware instructions.
 178  *
 179  * The BranchTarget field isn't needed, however.  Mesa's GLSL compiler
 180  * emits CAL and BGNSUB instructions with labels that can be used to map
 181  * subroutine calls to actual subroutine code blocks.
 182  *
 183  * The structures and function here implement patching of CAL instructions
 184  * so they jump to the right subroutine code...
 185  */
 186
 187
 188 /**
 189  * For each OPCODE_BGNSUB we create one of these.
 190  */
 191 struct brw_glsl_label
 192 {
 193    const char *name; /**< the label string */
 194    GLuint position;  /**< the position of the brw instruction for this label */
 195    struct brw_glsl_label *next;  /**< next in linked list */
 196 };
 197
 198
 199 /**
 200  * For each OPCODE_CAL we create one of these.
 201  */
 202 struct brw_glsl_call
 203 {
 204    GLuint call_inst_pos;  /**< location of the CAL instruction */
 205    const char *sub_name;  /**< name of subroutine to call */
 206    struct brw_glsl_call *next;  /**< next in linked list */
 207 };
 208
 209
 210 /**
 211  * Called for each OPCODE_BGNSUB.
 212  */
 213 void
 214 brw_save_label(struct brw_compile *c, const char *name, GLuint position)
 215 {
 216    struct brw_glsl_label *label = CALLOC_STRUCT(brw_glsl_label);
 217    label->name = name;
 218    label->position = position;
 219    label->next = c->first_label;
 220    c->first_label = label;
 221 }
 222
 223
 224 /**
 225  * Called for each OPCODE_CAL.
 226  */
 227 void
 228 brw_save_call(struct brw_compile *c, const char *name, GLuint call_pos)
 229 {
 230    struct brw_glsl_call *call = CALLOC_STRUCT(brw_glsl_call);
 231    call->call_inst_pos = call_pos;
 232    call->sub_name = name;
 233    call->next = c->first_call;
 234    c->first_call = call;
 235 }
 236
 237
 238 /**
 239  * Lookup a label, return label's position/offset.
 240  */
 241 static GLuint
 242 brw_lookup_label(struct brw_compile *c, const char *name)
 243 {
 244    const struct brw_glsl_label *label;
 245    for (label = c->first_label; label; label = label->next) {
 246       if (strcmp(name, label->name) == 0) {
 247          return label->position;
 248       }
 249    }
 250    abort();  /* should never happen */
 251    return ~0;
 252 }
 253
 254
 255 /**
 256  * When we're done generating code, this function is called to resolve
 257  * subroutine calls.
 258  */
 259 void
 260 brw_resolve_cals(struct brw_compile *c)
 261 {
 262     const struct brw_glsl_call *call;
 263
 264     for (call = c->first_call; call; call = call->next) {
 265         const GLuint sub_loc = brw_lookup_label(c, call->sub_name);
 266         struct brw_instruction *brw_call_inst = &c->store[call->call_inst_pos];
 267         struct brw_instruction *brw_sub_inst = &c->store[sub_loc];
 268         GLint offset = brw_sub_inst - brw_call_inst;
 269
 270         /* patch brw_inst1 to point to brw_inst2 */
 271         brw_set_src1(brw_call_inst, brw_imm_d(offset * 16));
 272     }
 273
 274     /* free linked list of calls */
 275     {
 276         struct brw_glsl_call *call, *next;
 277         for (call = c->first_call; call; call = next) {
 278             next = call->next;
 279             free(call);
 280         }
 281         c->first_call = NULL;
 282     }
 283
 284     /* free linked list of labels */
 285     {
 286         struct brw_glsl_label *label, *next;
 287         for (label = c->first_label; label; label = next) {
 288             next = label->next;
 289             free(label);
 290         }
 291         c->first_label = NULL;
 292     }
 293 }