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

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics 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 <keithw@vmware.com>
  30   */
  31
  32
  33
  34 #include "intel_batchbuffer.h"
  35 #include "brw_context.h"
  36 #include "brw_state.h"
  37 #include "brw_defines.h"
  38
  39 #define VS 0
  40 #define GS 1
  41 #define CLP 2
  42 #define SF 3
  43 #define CS 4
  44
  45 /** @file brw_urb.c
  46  *
  47  * Manages the division of the URB space between the various fixed-function
  48  * units.
  49  *
  50  * See the Thread Initiation Management section of the GEN4 B-Spec, and
  51  * the individual *_STATE structures for restrictions on numbers of
  52  * entries and threads.
  53  */
  54
  55 /*
  56  * Generally, a unit requires a min_nr_entries based on how many entries
  57  * it produces before the downstream unit gets unblocked and can use and
  58  * dereference some of its handles.
  59  *
  60  * The SF unit preallocates a PUE at the start of thread dispatch, and only
  61  * uses that one.  So it requires one entry per thread.
  62  *
  63  * For CLIP, the SF unit will hold the previous primitive while the
  64  * next is getting assembled, meaning that linestrips require 3 CLIP VUEs
  65  * (vertices) to ensure continued processing, trifans require 4, and tristrips
  66  * require 5.  There can be 1 or 2 threads, and each has the same requirement.
  67  *
  68  * GS has the same requirement as CLIP, but it never handles tristrips,
  69  * so we can lower the minimum to 4 for the POLYGONs (trifans) it produces.
  70  * We only run it single-threaded.
  71  *
  72  * For VS, the number of entries may be 8, 12, 16, or 32 (or 64 on G4X).
  73  * Each thread processes 2 preallocated VUEs (vertices) at a time, and they
  74  * get streamed down as soon as threads processing earlier vertices get
  75  * theirs accepted.
  76  *
  77  * Each unit will take the number of URB entries we give it (based on the
  78  * entry size calculated in brw_vs_emit.c for VUEs, brw_sf_emit.c for PUEs,
  79  * and brw_curbe.c for the CURBEs) and decide its maximum number of
  80  * threads it can support based on that. in brw_*_state.c.
  81  *
  82  * XXX: Are the min_entry_size numbers useful?
  83  * XXX: Verify min_nr_entries, esp for VS.
  84  * XXX: Verify SF min_entry_size.
  85  */
  86 static const struct {
  87    GLuint min_nr_entries;
  88    GLuint preferred_nr_entries;
  89    GLuint min_entry_size;
  90    GLuint max_entry_size;
  91 } limits[CS+1] = {
  92    { 16, 32, 1, 5 },                    /* vs */
  93    { 4, 8,  1, 5 },                     /* gs */
  94    { 5, 10,  1, 5 },                    /* clp */
  95    { 1, 8,  1, 12 },                    /* sf */
  96    { 1, 4,  1, 32 }                     /* cs */
  97 };
  98
  99
 100 static bool check_urb_layout(struct brw_context *brw)
 101 {
 102    brw->urb.vs_start = 0;
 103    brw->urb.gs_start = brw->urb.nr_vs_entries * brw->urb.vsize;
 104    brw->urb.clip_start = brw->urb.gs_start + brw->urb.nr_gs_entries * brw->urb.vsize;
 105    brw->urb.sf_start = brw->urb.clip_start + brw->urb.nr_clip_entries * brw->urb.vsize;
 106    brw->urb.cs_start = brw->urb.sf_start + brw->urb.nr_sf_entries * brw->urb.sfsize;
 107
 108    return brw->urb.cs_start + brw->urb.nr_cs_entries *
 109       brw->urb.csize <= brw->urb.size;
 110 }
 111
 112 /* Most minimal update, forces re-emit of URB fence packet after GS
 113  * unit turned on/off.
 114  */
 115 void
 116 brw_calculate_urb_fence(struct brw_context *brw, unsigned csize,
 117                         unsigned vsize, unsigned sfsize)
 118 {
 119    const struct gen_device_info *devinfo = &brw->screen->devinfo;
 120
 121    if (csize < limits[CS].min_entry_size)
 122       csize = limits[CS].min_entry_size;
 123
 124    if (vsize < limits[VS].min_entry_size)
 125       vsize = limits[VS].min_entry_size;
 126
 127    if (sfsize < limits[SF].min_entry_size)
 128       sfsize = limits[SF].min_entry_size;
 129
 130    if (brw->urb.vsize < vsize ||
 131        brw->urb.sfsize < sfsize ||
 132        brw->urb.csize < csize ||
 133        (brw->urb.constrained && (brw->urb.vsize > vsize ||
 134                                  brw->urb.sfsize > sfsize ||
 135                                  brw->urb.csize > csize))) {
 136
 137
 138       brw->urb.csize = csize;
 139       brw->urb.sfsize = sfsize;
 140       brw->urb.vsize = vsize;
 141
 142       brw->urb.nr_vs_entries = limits[VS].preferred_nr_entries;
 143       brw->urb.nr_gs_entries = limits[GS].preferred_nr_entries;
 144       brw->urb.nr_clip_entries = limits[CLP].preferred_nr_entries;
 145       brw->urb.nr_sf_entries = limits[SF].preferred_nr_entries;
 146       brw->urb.nr_cs_entries = limits[CS].preferred_nr_entries;
 147
 148       brw->urb.constrained = 0;
 149
 150       if (devinfo->gen == 5) {
 151          brw->urb.nr_vs_entries = 128;
 152          brw->urb.nr_sf_entries = 48;
 153          if (check_urb_layout(brw)) {
 154             goto done;
 155          } else {
 156             brw->urb.constrained = 1;
 157             brw->urb.nr_vs_entries = limits[VS].preferred_nr_entries;
 158             brw->urb.nr_sf_entries = limits[SF].preferred_nr_entries;
 159          }
 160       } else if (devinfo->is_g4x) {
 161          brw->urb.nr_vs_entries = 64;
 162          if (check_urb_layout(brw)) {
 163             goto done;
 164          } else {
 165             brw->urb.constrained = 1;
 166             brw->urb.nr_vs_entries = limits[VS].preferred_nr_entries;
 167          }
 168       }
 169
 170       if (!check_urb_layout(brw)) {
 171          brw->urb.nr_vs_entries = limits[VS].min_nr_entries;
 172          brw->urb.nr_gs_entries = limits[GS].min_nr_entries;
 173          brw->urb.nr_clip_entries = limits[CLP].min_nr_entries;
 174          brw->urb.nr_sf_entries = limits[SF].min_nr_entries;
 175          brw->urb.nr_cs_entries = limits[CS].min_nr_entries;
 176
 177          /* Mark us as operating with constrained nr_entries, so that next
 178           * time we recalculate we'll resize the fences in the hope of
 179           * escaping constrained mode and getting back to normal performance.
 180           */
 181          brw->urb.constrained = 1;
 182
 183          if (!check_urb_layout(brw)) {
 184             /* This is impossible, given the maximal sizes of urb
 185              * entries and the values for minimum nr of entries
 186              * provided above.
 187              */
 188             fprintf(stderr, "couldn't calculate URB layout!\n");
 189             exit(1);
 190          }
 191
 192          if (unlikely(INTEL_DEBUG & (DEBUG_URB|DEBUG_PERF)))
 193             fprintf(stderr, "URB CONSTRAINED\n");
 194       }
 195
 196 done:
 197       if (unlikely(INTEL_DEBUG & DEBUG_URB))
 198          fprintf(stderr,
 199                  "URB fence: %d ..VS.. %d ..GS.. %d ..CLP.. %d ..SF.. %d ..CS.. %d\n",
 200                  brw->urb.vs_start,
 201                  brw->urb.gs_start,
 202                  brw->urb.clip_start,
 203                  brw->urb.sf_start,
 204                  brw->urb.cs_start,
 205                  brw->urb.size);
 206
 207       brw->ctx.NewDriverState |= BRW_NEW_URB_FENCE;
 208    }
 209 }
 210
 211 static void recalculate_urb_fence( struct brw_context *brw )
 212 {
 213    brw_calculate_urb_fence(brw, brw->curbe.total_size,
 214                            brw_vue_prog_data(brw->vs.base.prog_data)->urb_entry_size,
 215                            brw->sf.prog_data->urb_entry_size);
 216 }
 217
 218
 219 const struct brw_tracked_state brw_recalculate_urb_fence = {
 220    .dirty = {
 221       .mesa = 0,
 222       .brw = BRW_NEW_BLORP |
 223              BRW_NEW_PUSH_CONSTANT_ALLOCATION |
 224              BRW_NEW_SF_PROG_DATA |
 225              BRW_NEW_VS_PROG_DATA,
 226    },
 227    .emit = recalculate_urb_fence
 228 };
 229
 230
 231
 232
 233
 234 void brw_upload_urb_fence(struct brw_context *brw)
 235 {
 236    struct brw_urb_fence uf;
 237    memset(&uf, 0, sizeof(uf));
 238
 239    uf.header.opcode = CMD_URB_FENCE;
 240    uf.header.length = sizeof(uf)/4-2;
 241    uf.header.vs_realloc = 1;
 242    uf.header.gs_realloc = 1;
 243    uf.header.clp_realloc = 1;
 244    uf.header.sf_realloc = 1;
 245    uf.header.vfe_realloc = 1;
 246    uf.header.cs_realloc = 1;
 247
 248    /* The ordering below is correct, not the layout in the
 249     * instruction.
 250     *
 251     * There are 256/384 urb reg pairs in total.
 252     */
 253    uf.bits0.vs_fence  = brw->urb.gs_start;
 254    uf.bits0.gs_fence  = brw->urb.clip_start;
 255    uf.bits0.clp_fence = brw->urb.sf_start;
 256    uf.bits1.sf_fence  = brw->urb.cs_start;
 257    uf.bits1.cs_fence  = brw->urb.size;
 258
 259    /* erratum: URB_FENCE must not cross a 64byte cacheline */
 260    if ((USED_BATCH(brw->batch) & 15) > 12) {
 261       int pad = 16 - (USED_BATCH(brw->batch) & 15);
 262       do
 263          *brw->batch.map_next++ = MI_NOOP;
 264       while (--pad);
 265    }
 266
 267    intel_batchbuffer_data(brw, &uf, sizeof(uf));
 268 }