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 "pipe/p_util.h"
  43
  44 #define FILE_DEBUG_FLAG DEBUG_FALLBACKS
  45
  46 /* Partition the CURBE between the various users of constant values:
  47  */
  48 static void calculate_curbe_offsets( struct brw_context *brw )
  49 {
  50    /* CACHE_NEW_WM_PROG */
  51    unsigned nr_fp_regs = align(brw->wm.prog_data->max_const, 16);
  52
  53    /* BRW_NEW_VERTEX_PROGRAM */
  54    unsigned nr_vp_regs = align(brw->vs.prog_data->max_const, 16);
  55    unsigned nr_clip_regs = 0;
  56    unsigned total_regs;
  57
  58 #if 0
  59    /* BRW_NEW_CLIP ? */
  60    if (brw->attribs.Transform->ClipPlanesEnabled) {
  61       unsigned nr_planes = 6 + brw_count_bits(brw->attribs.Transform->ClipPlanesEnabled);
  62       nr_clip_regs = align(nr_planes * 4, 16);
  63    }
  64 #endif
  65
  66
  67    total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs;
  68
  69    /* This can happen - what to do?  Probably rather than falling
  70     * back, the best thing to do is emit programs which code the
  71     * constants as immediate values.  Could do this either as a static
  72     * cap on WM and VS, or adaptively.
  73     *
  74     * Unfortunately, this is currently dependent on the results of the
  75     * program generation process (in the case of wm), so this would
  76     * introduce the need to re-generate programs in the event of a
  77     * curbe allocation failure.
  78     */
  79    /* Max size is 32 - just large enough to
  80     * hold the 128 parameters allowed by
  81     * the fragment and vertex program
  82     * api's.  It's not clear what happens
  83     * when both VP and FP want to use 128
  84     * parameters, though.
  85     */
  86    assert(total_regs <= 32);
  87
  88    /* Lazy resize:
  89     */
  90    if (nr_fp_regs > brw->curbe.wm_size ||
  91        nr_vp_regs > brw->curbe.vs_size ||
  92        nr_clip_regs != brw->curbe.clip_size ||
  93        (total_regs < brw->curbe.total_size / 4 &&
  94         brw->curbe.total_size > 16)) {
  95
  96       unsigned reg = 0;
  97
  98       /* Calculate a new layout:
  99        */
 100       reg = 0;
 101       brw->curbe.wm_start = reg;
 102       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
 103       brw->curbe.clip_start = reg;
 104       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
 105       brw->curbe.vs_start = reg;
 106       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
 107       brw->curbe.total_size = reg;
 108
 109 #if 0
 110       if (0)
 111          DBG("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 #endif
 119
 120       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 121    }
 122 }
 123
 124
 125 const struct brw_tracked_state brw_curbe_offsets = {
 126    .dirty = {
 127       .brw  = (BRW_NEW_CLIP |
 128                BRW_NEW_VS),
 129       .cache = CACHE_NEW_WM_PROG
 130    },
 131    .update = calculate_curbe_offsets
 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
 162 static float fixed_plane[6][4] = {
 163    { 0,    0,   -1, 1 },
 164    { 0,    0,    1, 1 },
 165    { 0,   -1,    0, 1 },
 166    { 0,    1,    0, 1 },
 167    {-1,    0,    0, 1 },
 168    { 1,    0,    0, 1 }
 169 };
 170
 171 /* Upload a new set of constants.  Too much variability to go into the
 172  * cache mechanism, but maybe would benefit from a comparison against
 173  * the current uploaded set of constants.
 174  */
 175 static void upload_constant_buffer(struct brw_context *brw)
 176 {
 177    struct brw_mem_pool *pool = &brw->pool[BRW_GS_POOL];
 178    unsigned sz = brw->curbe.total_size;
 179    unsigned bufsz = sz * sizeof(float);
 180    float *buf;
 181    unsigned i;
 182
 183
 184    if (sz == 0) {
 185       struct brw_constant_buffer cb;
 186       cb.header.opcode = CMD_CONST_BUFFER;
 187       cb.header.length = sizeof(cb)/4 - 2;
 188       cb.header.valid = 0;
 189       cb.bits0.buffer_length = 0;
 190       cb.bits0.buffer_address = 0;
 191       BRW_BATCH_STRUCT(brw, &cb);
 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 = (float *)malloc(bufsz);
 203
 204    memset(buf, 0, bufsz);
 205
 206    if (brw->curbe.wm_size) {
 207       unsigned offset = brw->curbe.wm_start * 16;
 208
 209       /* First the constant buffer constants:
 210        */
 211
 212       /* Then any internally generated constants:
 213        */
 214       for (i = 0; i < brw->wm.prog_data->nr_internal_consts; i++)
 215          buf[offset + i] = brw->wm.prog_data->internal_const[i];
 216
 217       assert(brw->wm.prog_data->max_const ==
 218              brw->wm.prog_data->nr_internal_consts);
 219    }
 220
 221
 222    /* The clipplanes are actually delivered to both CLIP and VS units.
 223     * VS uses them to calculate the outcode bitmasks.
 224     */
 225    if (brw->curbe.clip_size) {
 226       unsigned offset = brw->curbe.clip_start * 16;
 227       unsigned j;
 228
 229       /* If any planes are going this way, send them all this way:
 230        */
 231       for (i = 0; i < 6; i++) {
 232          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 233          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 234          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 235          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 236       }
 237
 238       /* Clip planes: BRW_NEW_CLIP:
 239        */
 240       for (j = 0; j < brw->attribs.Clip.nr; j++) {
 241          buf[offset + i * 4 + 0] = brw->attribs.Clip.ucp[j][0];
 242          buf[offset + i * 4 + 1] = brw->attribs.Clip.ucp[j][1];
 243          buf[offset + i * 4 + 2] = brw->attribs.Clip.ucp[j][2];
 244          buf[offset + i * 4 + 3] = brw->attribs.Clip.ucp[j][3];
 245          i++;
 246       }
 247    }
 248
 249
 250    if (brw->curbe.vs_size) {
 251       unsigned offset = brw->curbe.vs_start * 16;
 252       /*unsigned nr = vp->max_const;*/
 253       const struct pipe_constant_buffer *cbuffer = brw->attribs.Constants[0];
 254       struct pipe_winsys *ws = brw->pipe.winsys;
 255       /* FIXME: buffer size is num_consts + num_immediates */
 256       if (brw->vs.prog_data->num_consts) {
 257          /* map the vertex constant buffer and copy to curbe: */
 258          void *data = ws->buffer_map(ws, cbuffer->buffer, 0);
 259          /* FIXME: this is wrong. the cbuffer->size currently
 260           * represents size of consts + immediates. so if we'll
 261           * have both we'll copy over the end of the buffer
 262           * with the subsequent memcpy */
 263          memcpy(&buf[offset], data, cbuffer->size);
 264          ws->buffer_unmap(ws, cbuffer->buffer);
 265          offset += cbuffer->size;
 266       }
 267       /*immediates*/
 268       if (brw->vs.prog_data->num_imm) {
 269          memcpy(&buf[offset], brw->vs.prog_data->imm_buf,
 270                 brw->vs.prog_data->num_imm * 4 * sizeof(float));
 271       }
 272    }
 273
 274    if (1) {
 275       for (i = 0; i < sz; i+=4)
 276          debug_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 277                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 278
 279       debug_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 280                    brw->curbe.last_buf, buf,
 281                    bufsz, brw->curbe.last_bufsz,
 282                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 283    }
 284
 285    if (brw->curbe.last_buf &&
 286        bufsz == brw->curbe.last_bufsz &&
 287        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 288       free(buf);
 289 /*       return; */
 290    }
 291    else {
 292       if (brw->curbe.last_buf)
 293          free(brw->curbe.last_buf);
 294       brw->curbe.last_buf = buf;
 295       brw->curbe.last_bufsz = bufsz;
 296
 297
 298       if (!brw_pool_alloc(pool,
 299                           bufsz,
 300                           1 << 6,
 301                           &brw->curbe.gs_offset)) {
 302          debug_printf("out of GS memory for curbe\n");
 303          assert(0);
 304          return;
 305       }
 306
 307
 308       /* Copy data to the buffer:
 309        */
 310       brw->winsys->buffer_subdata_typed(brw->winsys,
 311                                         pool->buffer,
 312                                         brw->curbe.gs_offset,
 313                                         bufsz,
 314                                         buf,
 315                                         BRW_CONSTANT_BUFFER );
 316    }
 317
 318    /* TODO: only emit the constant_buffer packet when necessary, ie:
 319       - contents have changed
 320       - offset has changed
 321       - hw requirements due to other packets emitted.
 322    */
 323    {
 324       struct brw_constant_buffer cb;
 325
 326       memset(&cb, 0, sizeof(cb));
 327
 328       cb.header.opcode = CMD_CONST_BUFFER;
 329       cb.header.length = sizeof(cb)/4 - 2;
 330       cb.header.valid = 1;
 331       cb.bits0.buffer_length = sz - 1;
 332       cb.bits0.buffer_address = brw->curbe.gs_offset >> 6;
 333
 334       /* Because this provokes an action (ie copy the constants into the
 335        * URB), it shouldn't be shortcircuited if identical to the
 336        * previous time - because eg. the urb destination may have
 337        * changed, or the urb contents different to last time.
 338        *
 339        * Note that the data referred to is actually copied internally,
 340        * not just used in place according to passed pointer.
 341        *
 342        * It appears that the CS unit takes care of using each available
 343        * URB entry (Const URB Entry == CURBE) in turn, and issuing
 344        * flushes as necessary when doublebuffering of CURBEs isn't
 345        * possible.
 346        */
 347       BRW_BATCH_STRUCT(brw, &cb);
 348    }
 349 }
 350
 351 /* This tracked state is unique in that the state it monitors varies
 352  * dynamically depending on the parameters tracked by the fragment and
 353  * vertex programs.  This is the template used as a starting point,
 354  * each context will maintain a copy of this internally and update as
 355  * required.
 356  */
 357 const struct brw_tracked_state brw_constant_buffer = {
 358    .dirty = {
 359       .brw  = (BRW_NEW_CLIP |
 360                BRW_NEW_CONSTANTS |
 361                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 362                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 363                BRW_NEW_CURBE_OFFSETS),
 364       .cache = (CACHE_NEW_WM_PROG)
 365    },
 366    .update = upload_constant_buffer
 367 };
 368