src/gallium/drivers/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 #include "util/u_memory.h"
  33 #include "util/u_math.h"
  34
  35 #include "brw_batchbuffer.h"
  36 #include "brw_context.h"
  37 #include "brw_defines.h"
  38 #include "brw_state.h"
  39 #include "brw_debug.h"
  40
  41
  42 /**
  43  * Partition the CURBE between the various users of constant values:
  44  * Note that vertex and fragment shaders can now fetch constants out
  45  * of constant buffers.  We no longer allocatea block of the GRF for
  46  * constants.  That greatly reduces the demand for space in the CURBE.
  47  * Some of the comments within are dated...
  48  */
  49 static int calculate_curbe_offsets( struct brw_context *brw )
  50 {
  51    /* CACHE_NEW_WM_PROG */
  52    const GLuint nr_fp_regs = brw->wm.prog_data->curb_read_length;
  53
  54    /* BRW_NEW_VERTEX_PROGRAM */
  55    const GLuint nr_vp_regs = brw->vs.prog_data->curb_read_length;
  56    GLuint nr_clip_regs = 0;
  57    GLuint total_regs;
  58
  59    /* PIPE_NEW_CLIP */
  60    if (brw->curr.ucp.nr) {
  61       GLuint nr_planes = 6 + brw->curr.ucp.nr;
  62       nr_clip_regs = (nr_planes * 4 + 15) / 16;
  63    }
  64
  65
  66    total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs;
  67
  68    /* When this is > 32, want to use a true constant buffer to hold
  69     * the extra constants.
  70     */
  71    assert(total_regs <= 32);
  72
  73    /* Lazy resize:
  74     */
  75    if (nr_fp_regs > brw->curbe.wm_size ||
  76        nr_vp_regs > brw->curbe.vs_size ||
  77        nr_clip_regs != brw->curbe.clip_size ||
  78        (total_regs < brw->curbe.total_size / 4 &&
  79         brw->curbe.total_size > 16)) {
  80
  81       GLuint reg = 0;
  82
  83       /* Calculate a new layout:
  84        */
  85       reg = 0;
  86       brw->curbe.wm_start = reg;
  87       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
  88       brw->curbe.clip_start = reg;
  89       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
  90       brw->curbe.vs_start = reg;
  91       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
  92       brw->curbe.total_size = reg;
  93
  94       if (BRW_DEBUG & DEBUG_CURBE)
  95          debug_printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
  96                       brw->curbe.wm_start,
  97                       brw->curbe.wm_size,
  98                       brw->curbe.clip_start,
  99                       brw->curbe.clip_size,
 100                       brw->curbe.vs_start,
 101                       brw->curbe.vs_size );
 102
 103       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 104    }
 105
 106    return 0;
 107 }
 108
 109
 110 const struct brw_tracked_state brw_curbe_offsets = {
 111    .dirty = {
 112       .mesa = PIPE_NEW_CLIP,
 113       .brw  = BRW_NEW_VERTEX_PROGRAM,
 114       .cache = CACHE_NEW_WM_PROG
 115    },
 116    .prepare = calculate_curbe_offsets
 117 };
 118
 119
 120
 121
 122 /* Define the number of curbes within CS's urb allocation.  Multiple
 123  * urb entries -> multiple curbes.  These will be used by
 124  * fixed-function hardware in a double-buffering scheme to avoid a
 125  * pipeline stall each time the contents of the curbe is changed.
 126  */
 127 int brw_upload_cs_urb_state(struct brw_context *brw)
 128 {
 129    struct brw_cs_urb_state cs_urb;
 130    memset(&cs_urb, 0, sizeof(cs_urb));
 131
 132    /* It appears that this is the state packet for the CS unit, ie. the
 133     * urb entries detailed here are housed in the CS range from the
 134     * URB_FENCE command.
 135     */
 136    cs_urb.header.opcode = CMD_CS_URB_STATE;
 137    cs_urb.header.length = sizeof(cs_urb)/4 - 2;
 138
 139    /* BRW_NEW_URB_FENCE */
 140    cs_urb.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
 141    cs_urb.bits0.urb_entry_size = brw->urb.csize - 1;
 142
 143    assert(brw->urb.nr_cs_entries);
 144    BRW_CACHED_BATCH_STRUCT(brw, &cs_urb);
 145    return 0;
 146 }
 147
 148 static GLfloat fixed_plane[6][4] = {
 149    { 0,    0,   -1, 1 },
 150    { 0,    0,    1, 1 },
 151    { 0,   -1,    0, 1 },
 152    { 0,    1,    0, 1 },
 153    {-1,    0,    0, 1 },
 154    { 1,    0,    0, 1 }
 155 };
 156
 157 /* Upload a new set of constants.  Too much variability to go into the
 158  * cache mechanism, but maybe would benefit from a comparison against
 159  * the current uploaded set of constants.
 160  */
 161 static enum pipe_error prepare_curbe_buffer(struct brw_context *brw)
 162 {
 163    struct pipe_screen *screen = brw->base.screen;
 164    const GLuint sz = brw->curbe.total_size;
 165    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 166    enum pipe_error ret;
 167    GLfloat *buf;
 168    GLuint i;
 169
 170    if (sz == 0) {
 171       if (brw->curbe.last_buf) {
 172          free(brw->curbe.last_buf);
 173          brw->curbe.last_buf = NULL;
 174          brw->curbe.last_bufsz  = 0;
 175       }
 176       return 0;
 177    }
 178
 179    buf = (GLfloat *) CALLOC(bufsz, 1);
 180
 181    /* fragment shader constants */
 182    if (brw->curbe.wm_size) {
 183       const struct brw_fragment_shader *fs = brw->curr.fragment_shader;
 184       GLuint offset = brw->curbe.wm_start * 16;
 185       GLuint nr_immediate, nr_const;
 186
 187       nr_immediate = fs->immediates.nr;
 188       if (nr_immediate) {
 189          memcpy(&buf[offset],
 190                 fs->immediates.data,
 191                 nr_immediate * 4 * sizeof(float));
 192
 193          offset += nr_immediate * 4;
 194       }
 195
 196       nr_const = fs->info.file_max[TGSI_FILE_CONSTANT] + 1;
 197 /*      nr_const = brw->wm.prog_data->nr_params; */
 198       if (nr_const) {
 199          const GLfloat *value = screen->buffer_map( screen,
 200                                                     brw->curr.fragment_constants,
 201                                                     PIPE_BUFFER_USAGE_CPU_READ);
 202
 203          memcpy(&buf[offset], value,
 204                 nr_const * 4 * sizeof(float));
 205
 206          screen->buffer_unmap( screen,
 207                                brw->curr.fragment_constants );
 208       }
 209    }
 210
 211
 212    /* The clipplanes are actually delivered to both CLIP and VS units.
 213     * VS uses them to calculate the outcode bitmasks.
 214     */
 215    if (brw->curbe.clip_size) {
 216       GLuint offset = brw->curbe.clip_start * 16;
 217       GLuint j;
 218
 219       /* If any planes are going this way, send them all this way:
 220        */
 221       for (i = 0; i < 6; i++) {
 222          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 223          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 224          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 225          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 226       }
 227
 228       /* Clip planes:
 229        */
 230       assert(brw->curr.ucp.nr <= 6);
 231       for (j = 0; j < brw->curr.ucp.nr; j++) {
 232          buf[offset + i * 4 + 0] = brw->curr.ucp.ucp[j][0];
 233          buf[offset + i * 4 + 1] = brw->curr.ucp.ucp[j][1];
 234          buf[offset + i * 4 + 2] = brw->curr.ucp.ucp[j][2];
 235          buf[offset + i * 4 + 3] = brw->curr.ucp.ucp[j][3];
 236          i++;
 237       }
 238    }
 239
 240    /* vertex shader constants */
 241    if (brw->curbe.vs_size) {
 242       GLuint offset = brw->curbe.vs_start * 16;
 243       const struct brw_vertex_shader *vs = brw->curr.vertex_shader;
 244       GLuint nr_immediate, nr_const;
 245
 246       nr_immediate = vs->immediates.nr;
 247       if (nr_immediate) {
 248          memcpy(&buf[offset],
 249                 vs->immediates.data,
 250                 nr_immediate * 4 * sizeof(float));
 251
 252          offset += nr_immediate * 4;
 253       }
 254
 255       nr_const = vs->info.file_max[TGSI_FILE_CONSTANT] + 1;
 256       if (nr_const) {
 257          /* XXX: note that constant buffers are currently *already* in
 258           * buffer objects.  If we want to keep on putting them into the
 259           * curbe, makes sense to treat constbuf's specially with malloc.
 260           */
 261          const GLfloat *value = screen->buffer_map( screen,
 262                                                     brw->curr.vertex_constants,
 263                                                     PIPE_BUFFER_USAGE_CPU_READ);
 264
 265          /* XXX: what if user's constant buffer is too small?
 266           */
 267          memcpy(&buf[offset], value, nr_const * 4 * sizeof(float));
 268
 269          screen->buffer_unmap( screen, brw->curr.vertex_constants );
 270       }
 271    }
 272
 273    if (BRW_DEBUG & DEBUG_CURBE) {
 274       for (i = 0; i < sz*16; i+=4)
 275          debug_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 276                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 277
 278       debug_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 279                    (void *)brw->curbe.last_buf, (void *)buf,
 280                    bufsz, brw->curbe.last_bufsz,
 281                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 282    }
 283
 284    if (brw->curbe.curbe_bo != NULL &&
 285        brw->curbe.last_buf &&
 286        bufsz == brw->curbe.last_bufsz &&
 287        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 288       /* constants have not changed */
 289       FREE(buf);
 290    }
 291    else {
 292       /* constants have changed */
 293       FREE(brw->curbe.last_buf);
 294
 295       brw->curbe.last_buf = buf;
 296       brw->curbe.last_bufsz = bufsz;
 297
 298       if (brw->curbe.curbe_bo != NULL &&
 299           (brw->curbe.need_new_bo ||
 300            brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
 301       {
 302          bo_reference(&brw->curbe.curbe_bo, NULL);
 303       }
 304
 305       if (brw->curbe.curbe_bo == NULL) {
 306          /* Allocate a single page for CURBE entries for this
 307           * batchbuffer.  They're generally around 64b.  We will
 308           * discard the curbe buffer after the batch is flushed to
 309           * avoid synchronous updates.
 310           */
 311          ret = brw->sws->bo_alloc(brw->sws,
 312                                   BRW_BUFFER_TYPE_CURBE,
 313                                   4096, 1 << 6,
 314                                   &brw->curbe.curbe_bo);
 315          if (ret)
 316             return ret;
 317
 318          brw->curbe.curbe_next_offset = 0;
 319       }
 320
 321       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 322       brw->curbe.curbe_next_offset += bufsz;
 323       brw->curbe.curbe_next_offset = align(brw->curbe.curbe_next_offset, 64);
 324
 325       /* Copy data to the buffer:
 326        */
 327       brw->sws->bo_subdata(brw->curbe.curbe_bo,
 328                            BRW_DATA_CONSTANT_BUFFER,
 329                            brw->curbe.curbe_offset,
 330                            bufsz,
 331                            buf,
 332                            NULL, 0);
 333    }
 334
 335    brw_add_validated_bo(brw, brw->curbe.curbe_bo);
 336
 337    /* Because this provokes an action (ie copy the constants into the
 338     * URB), it shouldn't be shortcircuited if identical to the
 339     * previous time - because eg. the urb destination may have
 340     * changed, or the urb contents different to last time.
 341     *
 342     * Note that the data referred to is actually copied internally,
 343     * not just used in place according to passed pointer.
 344     *
 345     * It appears that the CS unit takes care of using each available
 346     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 347     * flushes as necessary when doublebuffering of CURBEs isn't
 348     * possible.
 349     */
 350
 351    return 0;
 352 }
 353
 354 static enum pipe_error emit_curbe_buffer(struct brw_context *brw)
 355 {
 356    GLuint sz = brw->curbe.total_size;
 357
 358    BEGIN_BATCH(2, IGNORE_CLIPRECTS);
 359    if (sz == 0) {
 360       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 361       OUT_BATCH(0);
 362    } else {
 363       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 364       OUT_RELOC(brw->curbe.curbe_bo,
 365                 BRW_USAGE_STATE,
 366                 (sz - 1) + brw->curbe.curbe_offset);
 367    }
 368    ADVANCE_BATCH();
 369    return 0;
 370 }
 371
 372 const struct brw_tracked_state brw_curbe_buffer = {
 373    .dirty = {
 374       .mesa = (PIPE_NEW_FRAGMENT_CONSTANTS |
 375                PIPE_NEW_VERTEX_CONSTANTS |
 376                PIPE_NEW_CLIP),
 377       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 378                BRW_NEW_VERTEX_PROGRAM |
 379                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 380                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 381                BRW_NEW_CURBE_OFFSETS |
 382                BRW_NEW_BATCH),
 383       .cache = (CACHE_NEW_WM_PROG)
 384    },
 385    .prepare = prepare_curbe_buffer,
 386    .emit = emit_curbe_buffer,
 387 };
 388