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 "glheader.h"
  35 #include "context.h"
  36 #include "macros.h"
  37 #include "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 #include "brw_aub.h"
  46
  47
  48 /* Partition the CURBE between the various users of constant values:
  49  */
  50 static void calculate_curbe_offsets( struct brw_context *brw )
  51 {
  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 (brw->attribs.Transform->ClipPlanesEnabled) {
  63       GLuint nr_planes = 6 + brw_count_bits(brw->attribs.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    .update = 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 #if 0
 162 const struct brw_tracked_state brw_constant_buffer_state = {
 163    .dirty = {
 164       .mesa = 0,
 165       .brw = BRW_NEW_URB_FENCE,
 166       .cache = 0
 167    },
 168    .update = brw_upload_constant_buffer_state
 169 };
 170 #endif
 171
 172
 173 static GLfloat fixed_plane[6][4] = {
 174    { 0,    0,   -1, 1 },
 175    { 0,    0,    1, 1 },
 176    { 0,   -1,    0, 1 },
 177    { 0,    1,    0, 1 },
 178    {-1,    0,    0, 1 },
 179    { 1,    0,    0, 1 }
 180 };
 181
 182 /* Upload a new set of constants.  Too much variability to go into the
 183  * cache mechanism, but maybe would benefit from a comparison against
 184  * the current uploaded set of constants.
 185  */
 186 static void upload_constant_buffer(struct brw_context *brw)
 187 {
 188    GLcontext *ctx = &brw->intel.ctx;
 189    struct brw_vertex_program *vp = (struct brw_vertex_program *)brw->vertex_program;
 190    struct brw_fragment_program *fp = (struct brw_fragment_program *)brw->fragment_program;
 191    struct brw_mem_pool *pool = &brw->pool[BRW_GS_POOL];
 192    GLuint sz = brw->curbe.total_size;
 193    GLuint bufsz = sz * 16 * sizeof(GLfloat);
 194    GLfloat *buf;
 195    GLuint i;
 196
 197    /* Update our own dependency flags.  This works because this
 198     * function will also be called whenever fp or vp changes.
 199     */
 200    brw->curbe.tracked_state.dirty.mesa = (_NEW_TRANSFORM|_NEW_PROJECTION);
 201    brw->curbe.tracked_state.dirty.mesa |= vp->param_state;
 202    brw->curbe.tracked_state.dirty.mesa |= fp->param_state;
 203
 204    if (sz == 0) {
 205       struct brw_constant_buffer cb;
 206       cb.header.opcode = CMD_CONST_BUFFER;
 207       cb.header.length = sizeof(cb)/4 - 2;
 208       cb.header.valid = 0;
 209       cb.bits0.buffer_length = 0;
 210       cb.bits0.buffer_address = 0;
 211       BRW_BATCH_STRUCT(brw, &cb);
 212
 213       if (brw->curbe.last_buf) {
 214          free(brw->curbe.last_buf);
 215          brw->curbe.last_buf = NULL;
 216          brw->curbe.last_bufsz  = 0;
 217       }
 218
 219       return;
 220    }
 221
 222    buf = (GLfloat *)malloc(bufsz);
 223
 224    memset(buf, 0, bufsz);
 225
 226    if (brw->curbe.wm_size) {
 227       GLuint offset = brw->curbe.wm_start * 16;
 228
 229       _mesa_load_state_parameters(ctx, fp->program.Base.Parameters);
 230
 231       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 232          buf[offset + i] = brw->wm.prog_data->param[i][0];
 233    }
 234
 235
 236    /* The clipplanes are actually delivered to both CLIP and VS units.
 237     * VS uses them to calculate the outcode bitmasks.
 238     */
 239    if (brw->curbe.clip_size) {
 240       GLuint offset = brw->curbe.clip_start * 16;
 241       GLuint j;
 242
 243       /* If any planes are going this way, send them all this way:
 244        */
 245       for (i = 0; i < 6; i++) {
 246          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 247          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 248          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 249          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 250       }
 251
 252       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 253        * clip-space:
 254        */
 255       assert(MAX_CLIP_PLANES == 6);
 256       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 257          if (brw->attribs.Transform->ClipPlanesEnabled & (1<<j)) {
 258             buf[offset + i * 4 + 0] = brw->attribs.Transform->_ClipUserPlane[j][0];
 259             buf[offset + i * 4 + 1] = brw->attribs.Transform->_ClipUserPlane[j][1];
 260             buf[offset + i * 4 + 2] = brw->attribs.Transform->_ClipUserPlane[j][2];
 261             buf[offset + i * 4 + 3] = brw->attribs.Transform->_ClipUserPlane[j][3];
 262             i++;
 263          }
 264       }
 265    }
 266
 267
 268    if (brw->curbe.vs_size) {
 269       GLuint offset = brw->curbe.vs_start * 16;
 270       GLuint nr = vp->program.Base.Parameters->NumParameters;
 271
 272       _mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
 273
 274       for (i = 0; i < nr; i++) {
 275          buf[offset + i * 4 + 0] = vp->program.Base.Parameters->ParameterValues[i][0];
 276          buf[offset + i * 4 + 1] = vp->program.Base.Parameters->ParameterValues[i][1];
 277          buf[offset + i * 4 + 2] = vp->program.Base.Parameters->ParameterValues[i][2];
 278          buf[offset + i * 4 + 3] = vp->program.Base.Parameters->ParameterValues[i][3];
 279       }
 280    }
 281
 282    if (0) {
 283       for (i = 0; i < sz*16; i+=4)
 284          _mesa_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 285                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 286
 287       _mesa_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 288                    brw->curbe.last_buf, buf,
 289                    bufsz, brw->curbe.last_bufsz,
 290                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 291    }
 292
 293    if (brw->curbe.last_buf &&
 294        bufsz == brw->curbe.last_bufsz &&
 295        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 296       free(buf);
 297 /*       return; */
 298    }
 299    else {
 300       if (brw->curbe.last_buf)
 301          free(brw->curbe.last_buf);
 302       brw->curbe.last_buf = buf;
 303       brw->curbe.last_bufsz = bufsz;
 304
 305
 306       if (!brw_pool_alloc(pool,
 307                           bufsz,
 308                           6,
 309                           &brw->curbe.gs_offset)) {
 310          _mesa_printf("out of GS memory for curbe\n");
 311          assert(0);
 312          return;
 313       }
 314
 315
 316       /* Copy data to the buffer:
 317        */
 318       bmBufferSubDataAUB(&brw->intel,
 319                          pool->buffer,
 320                          brw->curbe.gs_offset,
 321                          bufsz,
 322                          buf,
 323                          DW_CONSTANT_BUFFER,
 324                          0);
 325    }
 326
 327    /* TODO: only emit the constant_buffer packet when necessary, ie:
 328       - contents have changed
 329       - offset has changed
 330       - hw requirements due to other packets emitted.
 331    */
 332    {
 333       struct brw_constant_buffer cb;
 334
 335       memset(&cb, 0, sizeof(cb));
 336
 337       cb.header.opcode = CMD_CONST_BUFFER;
 338       cb.header.length = sizeof(cb)/4 - 2;
 339       cb.header.valid = 1;
 340       cb.bits0.buffer_length = sz - 1;
 341       cb.bits0.buffer_address = brw->curbe.gs_offset >> 6;
 342
 343       /* Because this provokes an action (ie copy the constants into the
 344        * URB), it shouldn't be shortcircuited if identical to the
 345        * previous time - because eg. the urb destination may have
 346        * changed, or the urb contents different to last time.
 347        *
 348        * Note that the data referred to is actually copied internally,
 349        * not just used in place according to passed pointer.
 350        *
 351        * It appears that the CS unit takes care of using each available
 352        * URB entry (Const URB Entry == CURBE) in turn, and issuing
 353        * flushes as necessary when doublebuffering of CURBEs isn't
 354        * possible.
 355        */
 356 /*       intel_batchbuffer_align(brw->intel.batch, 64, sizeof(cb)); */
 357       BRW_BATCH_STRUCT(brw, &cb);
 358 /*       intel_batchbuffer_align(brw->intel.batch, 64, 0); */
 359    }
 360 }
 361
 362 /* This tracked state is unique in that the state it monitors varies
 363  * dynamically depending on the parameters tracked by the fragment and
 364  * vertex programs.  This is the template used as a starting point,
 365  * each context will maintain a copy of this internally and update as
 366  * required.
 367  */
 368 const struct brw_tracked_state brw_constant_buffer = {
 369    .dirty = {
 370       .mesa = (_NEW_TRANSFORM|_NEW_PROJECTION),      /* plus fp and vp flags */
 371       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 372                BRW_NEW_VERTEX_PROGRAM |
 373                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 374                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 375                BRW_NEW_CURBE_OFFSETS),
 376       .cache = (CACHE_NEW_WM_PROG)
 377    },
 378    .update = upload_constant_buffer
 379 };
 380