src/gallium/drivers/i965simple/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 "brw_context.h"
  35 #include "brw_defines.h"
  36 #include "brw_state.h"
  37 #include "brw_batch.h"
  38 #include "brw_util.h"
  39 #include "brw_wm.h"
  40 #include "pipe/p_state.h"
  41 #include "pipe/p_winsys.h"
  42 #include "util/u_math.h"
  43 #include "util/u_memory.h"
  44
  45 #define FILE_DEBUG_FLAG DEBUG_FALLBACKS
  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    /* CACHE_NEW_WM_PROG */
  52    unsigned nr_fp_regs = align(brw->wm.prog_data->max_const, 16);
  53
  54    /* BRW_NEW_VERTEX_PROGRAM */
  55    unsigned nr_vp_regs = align(brw->vs.prog_data->max_const, 16);
  56    unsigned nr_clip_regs = 0;
  57    unsigned total_regs;
  58
  59 #if 0
  60    /* BRW_NEW_CLIP ? */
  61    if (brw->attribs.Transform->ClipPlanesEnabled) {
  62       unsigned nr_planes = 6 + brw_count_bits(brw->attribs.Transform->ClipPlanesEnabled);
  63       nr_clip_regs = align(nr_planes * 4, 16);
  64    }
  65 #endif
  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       unsigned 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       if (0)
 112          DBG("curbe wm %d+%d clip %d+%d vs %d+%d\n",
 113                       brw->curbe.wm_start,
 114                       brw->curbe.wm_size,
 115                       brw->curbe.clip_start,
 116                       brw->curbe.clip_size,
 117                       brw->curbe.vs_start,
 118                       brw->curbe.vs_size );
 119 #endif
 120
 121       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 122    }
 123 }
 124
 125
 126 const struct brw_tracked_state brw_curbe_offsets = {
 127    .dirty = {
 128       .brw  = (BRW_NEW_CLIP |
 129                BRW_NEW_VS),
 130       .cache = CACHE_NEW_WM_PROG
 131    },
 132    .update = calculate_curbe_offsets
 133 };
 134
 135
 136
 137 /* Define the number of curbes within CS's urb allocation.  Multiple
 138  * urb entries -> multiple curbes.  These will be used by
 139  * fixed-function hardware in a double-buffering scheme to avoid a
 140  * pipeline stall each time the contents of the curbe is changed.
 141  */
 142 void brw_upload_constant_buffer_state(struct brw_context *brw)
 143 {
 144    struct brw_constant_buffer_state cbs;
 145    memset(&cbs, 0, sizeof(cbs));
 146
 147    /* It appears that this is the state packet for the CS unit, ie. the
 148     * urb entries detailed here are housed in the CS range from the
 149     * URB_FENCE command.
 150     */
 151    cbs.header.opcode = CMD_CONST_BUFFER_STATE;
 152    cbs.header.length = sizeof(cbs)/4 - 2;
 153
 154    /* BRW_NEW_URB_FENCE */
 155    cbs.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
 156    cbs.bits0.urb_entry_size = brw->urb.csize - 1;
 157
 158    assert(brw->urb.nr_cs_entries);
 159    BRW_CACHED_BATCH_STRUCT(brw, &cbs);
 160 }
 161
 162
 163 static float fixed_plane[6][4] = {
 164    { 0,    0,   -1, 1 },
 165    { 0,    0,    1, 1 },
 166    { 0,   -1,    0, 1 },
 167    { 0,    1,    0, 1 },
 168    {-1,    0,    0, 1 },
 169    { 1,    0,    0, 1 }
 170 };
 171
 172 /* Upload a new set of constants.  Too much variability to go into the
 173  * cache mechanism, but maybe would benefit from a comparison against
 174  * the current uploaded set of constants.
 175  */
 176 static void upload_constant_buffer(struct brw_context *brw)
 177 {
 178    struct brw_mem_pool *pool = &brw->pool[BRW_GS_POOL];
 179    unsigned sz = brw->curbe.total_size;
 180    unsigned bufsz = sz * sizeof(float);
 181    float *buf;
 182    unsigned i;
 183
 184
 185    if (sz == 0) {
 186       struct brw_constant_buffer cb;
 187       cb.header.opcode = CMD_CONST_BUFFER;
 188       cb.header.length = sizeof(cb)/4 - 2;
 189       cb.header.valid = 0;
 190       cb.bits0.buffer_length = 0;
 191       cb.bits0.buffer_address = 0;
 192       BRW_BATCH_STRUCT(brw, &cb);
 193
 194       if (brw->curbe.last_buf) {
 195          free(brw->curbe.last_buf);
 196          brw->curbe.last_buf = NULL;
 197          brw->curbe.last_bufsz  = 0;
 198       }
 199
 200       return;
 201    }
 202
 203    buf = (float *)malloc(bufsz);
 204
 205    memset(buf, 0, bufsz);
 206
 207    if (brw->curbe.wm_size) {
 208       unsigned offset = brw->curbe.wm_start * 16;
 209
 210       /* First the constant buffer constants:
 211        */
 212
 213       /* Then any internally generated constants:
 214        */
 215       for (i = 0; i < brw->wm.prog_data->nr_internal_consts; i++)
 216          buf[offset + i] = brw->wm.prog_data->internal_const[i];
 217
 218       assert(brw->wm.prog_data->max_const ==
 219              brw->wm.prog_data->nr_internal_consts);
 220    }
 221
 222
 223    /* The clipplanes are actually delivered to both CLIP and VS units.
 224     * VS uses them to calculate the outcode bitmasks.
 225     */
 226    if (brw->curbe.clip_size) {
 227       unsigned offset = brw->curbe.clip_start * 16;
 228       unsigned j;
 229
 230       /* If any planes are going this way, send them all this way:
 231        */
 232       for (i = 0; i < 6; i++) {
 233          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 234          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 235          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 236          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 237       }
 238
 239       /* Clip planes: BRW_NEW_CLIP:
 240        */
 241       for (j = 0; j < brw->attribs.Clip.nr; j++) {
 242          buf[offset + i * 4 + 0] = brw->attribs.Clip.ucp[j][0];
 243          buf[offset + i * 4 + 1] = brw->attribs.Clip.ucp[j][1];
 244          buf[offset + i * 4 + 2] = brw->attribs.Clip.ucp[j][2];
 245          buf[offset + i * 4 + 3] = brw->attribs.Clip.ucp[j][3];
 246          i++;
 247       }
 248    }
 249
 250
 251    if (brw->curbe.vs_size) {
 252       unsigned offset = brw->curbe.vs_start * 16;
 253       /*unsigned nr = vp->max_const;*/
 254       const struct pipe_constant_buffer *cbuffer = brw->attribs.Constants[0];
 255       struct pipe_winsys *ws = brw->pipe.winsys;
 256       /* FIXME: buffer size is num_consts + num_immediates */
 257       if (brw->vs.prog_data->num_consts) {
 258          /* map the vertex constant buffer and copy to curbe: */
 259          void *data = ws->buffer_map(ws, cbuffer->buffer, 0);
 260          /* FIXME: this is wrong. the cbuffer->size currently
 261           * represents size of consts + immediates. so if we'll
 262           * have both we'll copy over the end of the buffer
 263           * with the subsequent memcpy */
 264          memcpy(&buf[offset], data, cbuffer->size);
 265          ws->buffer_unmap(ws, cbuffer->buffer);
 266          offset += cbuffer->size;
 267       }
 268       /*immediates*/
 269       if (brw->vs.prog_data->num_imm) {
 270          memcpy(&buf[offset], brw->vs.prog_data->imm_buf,
 271                 brw->vs.prog_data->num_imm * 4 * sizeof(float));
 272       }
 273    }
 274
 275    if (1) {
 276       for (i = 0; i < sz; i+=4)
 277          debug_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 278                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 279
 280       debug_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 281                    brw->curbe.last_buf, buf,
 282                    bufsz, brw->curbe.last_bufsz,
 283                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 284    }
 285
 286    if (brw->curbe.last_buf &&
 287        bufsz == brw->curbe.last_bufsz &&
 288        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 289       free(buf);
 290 /*       return; */
 291    }
 292    else {
 293       if (brw->curbe.last_buf)
 294          free(brw->curbe.last_buf);
 295       brw->curbe.last_buf = buf;
 296       brw->curbe.last_bufsz = bufsz;
 297
 298
 299       if (!brw_pool_alloc(pool,
 300                           bufsz,
 301                           1 << 6,
 302                           &brw->curbe.gs_offset)) {
 303          debug_printf("out of GS memory for curbe\n");
 304          assert(0);
 305          return;
 306       }
 307
 308
 309       /* Copy data to the buffer:
 310        */
 311       brw->winsys->buffer_subdata_typed(brw->winsys,
 312                                         pool->buffer,
 313                                         brw->curbe.gs_offset,
 314                                         bufsz,
 315                                         buf,
 316                                         BRW_CONSTANT_BUFFER );
 317    }
 318
 319    /* TODO: only emit the constant_buffer packet when necessary, ie:
 320       - contents have changed
 321       - offset has changed
 322       - hw requirements due to other packets emitted.
 323    */
 324    {
 325       struct brw_constant_buffer cb;
 326
 327       memset(&cb, 0, sizeof(cb));
 328
 329       cb.header.opcode = CMD_CONST_BUFFER;
 330       cb.header.length = sizeof(cb)/4 - 2;
 331       cb.header.valid = 1;
 332       cb.bits0.buffer_length = sz - 1;
 333       cb.bits0.buffer_address = brw->curbe.gs_offset >> 6;
 334
 335       /* Because this provokes an action (ie copy the constants into the
 336        * URB), it shouldn't be shortcircuited if identical to the
 337        * previous time - because eg. the urb destination may have
 338        * changed, or the urb contents different to last time.
 339        *
 340        * Note that the data referred to is actually copied internally,
 341        * not just used in place according to passed pointer.
 342        *
 343        * It appears that the CS unit takes care of using each available
 344        * URB entry (Const URB Entry == CURBE) in turn, and issuing
 345        * flushes as necessary when doublebuffering of CURBEs isn't
 346        * possible.
 347        */
 348       BRW_BATCH_STRUCT(brw, &cb);
 349    }
 350 }
 351
 352 /* This tracked state is unique in that the state it monitors varies
 353  * dynamically depending on the parameters tracked by the fragment and
 354  * vertex programs.  This is the template used as a starting point,
 355  * each context will maintain a copy of this internally and update as
 356  * required.
 357  */
 358 const struct brw_tracked_state brw_constant_buffer = {
 359    .dirty = {
 360       .brw  = (BRW_NEW_CLIP |
 361                BRW_NEW_CONSTANTS |
 362                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 363                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 364                BRW_NEW_CURBE_OFFSETS),
 365       .cache = (CACHE_NEW_WM_PROG)
 366    },
 367    .update = upload_constant_buffer
 368 };
 369