src/mesa/drivers/dri/i965/brw_curbe.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
  34 #include "main/glheader.h"
  35 #include "main/context.h"
  36 #include "main/macros.h"
  37 #include "main/enums.h"
  38 #include "shader/prog_parameter.h"
  39 #include "shader/prog_statevars.h"
  40 #include "intel_batchbuffer.h"
  41 #include "brw_context.h"
  42 #include "brw_defines.h"
  43 #include "brw_state.h"
  44 #include "brw_util.h"
  45
  46
  47 /* Partition the CURBE between the various users of constant values:
  48  */
  49 static void calculate_curbe_offsets( struct brw_context *brw )
  50 {
  51    GLcontext *ctx = &brw->intel.ctx;
  52    /* CACHE_NEW_WM_PROG */
  53    GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16;
  54
  55    /* BRW_NEW_VERTEX_PROGRAM */
  56    struct brw_vertex_program *vp = (struct brw_vertex_program *)brw->vertex_program;
  57    GLuint nr_vp_regs = (vp->program.Base.Parameters->NumParameters * 4 + 15) / 16;
  58    GLuint nr_clip_regs = 0;
  59    GLuint total_regs;
  60
  61    /* _NEW_TRANSFORM */
  62    if (ctx->Transform.ClipPlanesEnabled) {
  63       GLuint nr_planes = 6 + brw_count_bits(ctx->Transform.ClipPlanesEnabled);
  64       nr_clip_regs = (nr_planes * 4 + 15) / 16;
  65    }
  66
  67
  68    total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs;
  69
  70    /* This can happen - what to do?  Probably rather than falling
  71     * back, the best thing to do is emit programs which code the
  72     * constants as immediate values.  Could do this either as a static
  73     * cap on WM and VS, or adaptively.
  74     *
  75     * Unfortunately, this is currently dependent on the results of the
  76     * program generation process (in the case of wm), so this would
  77     * introduce the need to re-generate programs in the event of a
  78     * curbe allocation failure.
  79     */
  80    /* Max size is 32 - just large enough to
  81     * hold the 128 parameters allowed by
  82     * the fragment and vertex program
  83     * api's.  It's not clear what happens
  84     * when both VP and FP want to use 128
  85     * parameters, though.
  86     */
  87    assert(total_regs <= 32);
  88
  89    /* Lazy resize:
  90     */
  91    if (nr_fp_regs > brw->curbe.wm_size ||
  92        nr_vp_regs > brw->curbe.vs_size ||
  93        nr_clip_regs != brw->curbe.clip_size ||
  94        (total_regs < brw->curbe.total_size / 4 &&
  95         brw->curbe.total_size > 16)) {
  96
  97       GLuint reg = 0;
  98
  99       /* Calculate a new layout:
 100        */
 101       reg = 0;
 102       brw->curbe.wm_start = reg;
 103       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
 104       brw->curbe.clip_start = reg;
 105       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
 106       brw->curbe.vs_start = reg;
 107       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
 108       brw->curbe.total_size = reg;
 109
 110       if (0)
 111          _mesa_printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
 112                       brw->curbe.wm_start,
 113                       brw->curbe.wm_size,
 114                       brw->curbe.clip_start,
 115                       brw->curbe.clip_size,
 116                       brw->curbe.vs_start,
 117                       brw->curbe.vs_size );
 118
 119       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 120    }
 121 }
 122
 123
 124 const struct brw_tracked_state brw_curbe_offsets = {
 125    .dirty = {
 126       .mesa = _NEW_TRANSFORM,
 127       .brw  = BRW_NEW_VERTEX_PROGRAM,
 128       .cache = CACHE_NEW_WM_PROG
 129    },
 130    .prepare = calculate_curbe_offsets
 131 };
 132
 133
 134
 135
 136 /* Define the number of curbes within CS's urb allocation.  Multiple
 137  * urb entries -> multiple curbes.  These will be used by
 138  * fixed-function hardware in a double-buffering scheme to avoid a
 139  * pipeline stall each time the contents of the curbe is changed.
 140  */
 141 void brw_upload_constant_buffer_state(struct brw_context *brw)
 142 {
 143    struct brw_constant_buffer_state cbs;
 144    memset(&cbs, 0, sizeof(cbs));
 145
 146    /* It appears that this is the state packet for the CS unit, ie. the
 147     * urb entries detailed here are housed in the CS range from the
 148     * URB_FENCE command.
 149     */
 150    cbs.header.opcode = CMD_CONST_BUFFER_STATE;
 151    cbs.header.length = sizeof(cbs)/4 - 2;
 152
 153    /* BRW_NEW_URB_FENCE */
 154    cbs.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
 155    cbs.bits0.urb_entry_size = brw->urb.csize - 1;
 156
 157    assert(brw->urb.nr_cs_entries);
 158    BRW_CACHED_BATCH_STRUCT(brw, &cbs);
 159 }
 160
 161 static GLfloat fixed_plane[6][4] = {
 162    { 0,    0,   -1, 1 },
 163    { 0,    0,    1, 1 },
 164    { 0,   -1,    0, 1 },
 165    { 0,    1,    0, 1 },
 166    {-1,    0,    0, 1 },
 167    { 1,    0,    0, 1 }
 168 };
 169
 170 /* Upload a new set of constants.  Too much variability to go into the
 171  * cache mechanism, but maybe would benefit from a comparison against
 172  * the current uploaded set of constants.
 173  */
 174 static void prepare_constant_buffer(struct brw_context *brw)
 175 {
 176    GLcontext *ctx = &brw->intel.ctx;
 177    struct brw_vertex_program *vp = (struct brw_vertex_program *)brw->vertex_program;
 178    struct brw_fragment_program *fp = (struct brw_fragment_program *)brw->fragment_program;
 179    GLuint sz = brw->curbe.total_size;
 180    GLuint bufsz = sz * 16 * sizeof(GLfloat);
 181    GLfloat *buf;
 182    GLuint i;
 183
 184    /* Update our own dependency flags.  This works because this
 185     * function will also be called whenever fp or vp changes.
 186     */
 187    brw->curbe.tracked_state.dirty.mesa = (_NEW_TRANSFORM|_NEW_PROJECTION);
 188    brw->curbe.tracked_state.dirty.mesa |= vp->program.Base.Parameters->StateFlags;
 189    brw->curbe.tracked_state.dirty.mesa |= fp->program.Base.Parameters->StateFlags;
 190
 191    if (sz == 0) {
 192
 193       if (brw->curbe.last_buf) {
 194          free(brw->curbe.last_buf);
 195          brw->curbe.last_buf = NULL;
 196          brw->curbe.last_bufsz  = 0;
 197       }
 198
 199       return;
 200    }
 201
 202    buf = (GLfloat *)malloc(bufsz);
 203
 204    memset(buf, 0, bufsz);
 205
 206    if (brw->curbe.wm_size) {
 207       GLuint offset = brw->curbe.wm_start * 16;
 208
 209       _mesa_load_state_parameters(ctx, fp->program.Base.Parameters);
 210
 211       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 212          buf[offset + i] = brw->wm.prog_data->param[i][0];
 213    }
 214
 215
 216    /* The clipplanes are actually delivered to both CLIP and VS units.
 217     * VS uses them to calculate the outcode bitmasks.
 218     */
 219    if (brw->curbe.clip_size) {
 220       GLuint offset = brw->curbe.clip_start * 16;
 221       GLuint j;
 222
 223       /* If any planes are going this way, send them all this way:
 224        */
 225       for (i = 0; i < 6; i++) {
 226          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 227          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 228          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 229          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 230       }
 231
 232       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 233        * clip-space:
 234        */
 235       assert(MAX_CLIP_PLANES == 6);
 236       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 237          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 238             buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0];
 239             buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1];
 240             buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2];
 241             buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3];
 242             i++;
 243          }
 244       }
 245    }
 246
 247
 248    if (brw->curbe.vs_size) {
 249       GLuint offset = brw->curbe.vs_start * 16;
 250       GLuint nr = vp->program.Base.Parameters->NumParameters;
 251
 252       _mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
 253
 254       for (i = 0; i < nr; i++) {
 255          buf[offset + i * 4 + 0] = vp->program.Base.Parameters->ParameterValues[i][0];
 256          buf[offset + i * 4 + 1] = vp->program.Base.Parameters->ParameterValues[i][1];
 257          buf[offset + i * 4 + 2] = vp->program.Base.Parameters->ParameterValues[i][2];
 258          buf[offset + i * 4 + 3] = vp->program.Base.Parameters->ParameterValues[i][3];
 259       }
 260    }
 261
 262    if (0) {
 263       for (i = 0; i < sz*16; i+=4)
 264          _mesa_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 265                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 266
 267       _mesa_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 268                    brw->curbe.last_buf, buf,
 269                    bufsz, brw->curbe.last_bufsz,
 270                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 271    }
 272
 273    if (brw->curbe.curbe_bo != NULL &&
 274        brw->curbe.last_buf &&
 275        bufsz == brw->curbe.last_bufsz &&
 276        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 277       free(buf);
 278    }
 279    else {
 280       if (brw->curbe.last_buf)
 281          free(brw->curbe.last_buf);
 282       brw->curbe.last_buf = buf;
 283       brw->curbe.last_bufsz = bufsz;
 284
 285       if (brw->curbe.curbe_bo != NULL &&
 286           (brw->curbe.need_new_bo ||
 287            brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
 288       {
 289          dri_bo_unreference(brw->curbe.curbe_bo);
 290          brw->curbe.curbe_bo = NULL;
 291       }
 292
 293       if (brw->curbe.curbe_bo == NULL) {
 294          /* Allocate a single page for CURBE entries for this batchbuffer.
 295           * They're generally around 64b.
 296           */
 297          brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
 298                                             4096, 1 << 6);
 299          brw->curbe.curbe_next_offset = 0;
 300       }
 301
 302       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 303       brw->curbe.curbe_next_offset += bufsz;
 304       brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 305
 306       /* Copy data to the buffer:
 307        */
 308       dri_bo_subdata(brw->curbe.curbe_bo, brw->curbe.curbe_offset, bufsz, buf);
 309    }
 310
 311    brw_add_validated_bo(brw, brw->curbe.curbe_bo);
 312
 313    /* Because this provokes an action (ie copy the constants into the
 314     * URB), it shouldn't be shortcircuited if identical to the
 315     * previous time - because eg. the urb destination may have
 316     * changed, or the urb contents different to last time.
 317     *
 318     * Note that the data referred to is actually copied internally,
 319     * not just used in place according to passed pointer.
 320     *
 321     * It appears that the CS unit takes care of using each available
 322     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 323     * flushes as necessary when doublebuffering of CURBEs isn't
 324     * possible.
 325     */
 326 }
 327
 328
 329 static void emit_constant_buffer(struct brw_context *brw)
 330 {
 331    struct intel_context *intel = &brw->intel;
 332    GLuint sz = brw->curbe.total_size;
 333
 334    BEGIN_BATCH(2, IGNORE_CLIPRECTS);
 335    if (sz == 0) {
 336       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 337       OUT_BATCH(0);
 338    } else {
 339       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 340       OUT_RELOC(brw->curbe.curbe_bo,
 341                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 342                 (sz - 1) + brw->curbe.curbe_offset);
 343    }
 344    ADVANCE_BATCH();
 345 }
 346
 347 /* This tracked state is unique in that the state it monitors varies
 348  * dynamically depending on the parameters tracked by the fragment and
 349  * vertex programs.  This is the template used as a starting point,
 350  * each context will maintain a copy of this internally and update as
 351  * required.
 352  */
 353 const struct brw_tracked_state brw_constant_buffer = {
 354    .dirty = {
 355       .mesa = (_NEW_TRANSFORM|_NEW_PROJECTION),      /* plus fp and vp flags */
 356       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 357                BRW_NEW_VERTEX_PROGRAM |
 358                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 359                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 360                BRW_NEW_CURBE_OFFSETS |
 361                BRW_NEW_BATCH),
 362       .cache = (CACHE_NEW_WM_PROG)
 363    },
 364    .prepare = prepare_constant_buffer,
 365    .emit = emit_constant_buffer,
 366 };
 367