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_util.h"
  40 #include "brw_debug.h"
  41 #include "brw_screen.h"
  42
  43
  44 /**
  45  * Partition the CURBE between the various users of constant values:
  46  * Note that vertex and fragment shaders can now fetch constants out
  47  * of constant buffers.  We no longer allocatea block of the GRF for
  48  * constants.  That greatly reduces the demand for space in the CURBE.
  49  * Some of the comments within are dated...
  50  */
  51 static void calculate_curbe_offsets( struct brw_context *brw )
  52 {
  53    /* CACHE_NEW_WM_PROG */
  54    const GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16;
  55
  56    /* BRW_NEW_VERTEX_PROGRAM */
  57    const GLuint nr_vp_regs = (brw->vs.prog_data->nr_params + 15) / 16;
  58    GLuint nr_clip_regs = 0;
  59    GLuint total_regs;
  60
  61    /* PIPE_NEW_CLIP */
  62    if (brw->curr.ucp.nr) {
  63       GLuint nr_planes = 6 + brw->curr.ucp.nr;
  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    /* When this is > 32, want to use a true constant buffer to hold
  71     * the extra constants.
  72     */
  73    assert(total_regs <= 32);
  74
  75    /* Lazy resize:
  76     */
  77    if (nr_fp_regs > brw->curbe.wm_size ||
  78        nr_vp_regs > brw->curbe.vs_size ||
  79        nr_clip_regs != brw->curbe.clip_size ||
  80        (total_regs < brw->curbe.total_size / 4 &&
  81         brw->curbe.total_size > 16)) {
  82
  83       GLuint reg = 0;
  84
  85       /* Calculate a new layout:
  86        */
  87       reg = 0;
  88       brw->curbe.wm_start = reg;
  89       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
  90       brw->curbe.clip_start = reg;
  91       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
  92       brw->curbe.vs_start = reg;
  93       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
  94       brw->curbe.total_size = reg;
  95
  96       if (BRW_DEBUG & DEBUG_CURBE)
  97          debug_printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
  98                       brw->curbe.wm_start,
  99                       brw->curbe.wm_size,
 100                       brw->curbe.clip_start,
 101                       brw->curbe.clip_size,
 102                       brw->curbe.vs_start,
 103                       brw->curbe.vs_size );
 104
 105       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 106    }
 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 void 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 }
 146
 147 static GLfloat fixed_plane[6][4] = {
 148    { 0,    0,   -1, 1 },
 149    { 0,    0,    1, 1 },
 150    { 0,   -1,    0, 1 },
 151    { 0,    1,    0, 1 },
 152    {-1,    0,    0, 1 },
 153    { 1,    0,    0, 1 }
 154 };
 155
 156 /* Upload a new set of constants.  Too much variability to go into the
 157  * cache mechanism, but maybe would benefit from a comparison against
 158  * the current uploaded set of constants.
 159  */
 160 static void prepare_constant_buffer(struct brw_context *brw)
 161 {
 162    const GLuint sz = brw->curbe.total_size;
 163    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 164    GLfloat *buf;
 165    GLuint i;
 166
 167    if (sz == 0) {
 168       if (brw->curbe.last_buf) {
 169          free(brw->curbe.last_buf);
 170          brw->curbe.last_buf = NULL;
 171          brw->curbe.last_bufsz  = 0;
 172       }
 173       return;
 174    }
 175
 176    buf = (GLfloat *) CALLOC(bufsz, 1);
 177
 178    /* fragment shader constants */
 179    if (brw->curbe.wm_size) {
 180       GLuint offset = brw->curbe.wm_start * 16;
 181
 182       /* map fs constant buffer */
 183
 184       /* copy float constants */
 185       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 186          buf[offset + i] = *brw->wm.prog_data->param[i];
 187
 188       /* unmap fs constant buffer */
 189    }
 190
 191
 192    /* The clipplanes are actually delivered to both CLIP and VS units.
 193     * VS uses them to calculate the outcode bitmasks.
 194     */
 195    if (brw->curbe.clip_size) {
 196       GLuint offset = brw->curbe.clip_start * 16;
 197       GLuint j;
 198
 199       /* If any planes are going this way, send them all this way:
 200        */
 201       for (i = 0; i < 6; i++) {
 202          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 203          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 204          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 205          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 206       }
 207
 208       /* Clip planes:
 209        */
 210       assert(brw->curr.ucp.nr <= 6);
 211       for (j = 0; j < brw->curr.ucp.nr; j++) {
 212          buf[offset + i * 4 + 0] = brw->curr.ucp.ucp[j][0];
 213          buf[offset + i * 4 + 1] = brw->curr.ucp.ucp[j][1];
 214          buf[offset + i * 4 + 2] = brw->curr.ucp.ucp[j][2];
 215          buf[offset + i * 4 + 3] = brw->curr.ucp.ucp[j][3];
 216          i++;
 217       }
 218    }
 219
 220    /* vertex shader constants */
 221    if (brw->curbe.vs_size) {
 222       GLuint offset = brw->curbe.vs_start * 16;
 223       GLuint nr = brw->curr.vertex_shader->info.file_max[TGSI_FILE_CONSTANT];
 224       struct pipe_screen *screen = &brw->brw_screen->base;
 225
 226       const GLfloat *value = screen->buffer_map( screen,
 227                                                  brw->curr.vertex_constants,
 228                                                  PIPE_BUFFER_USAGE_CPU_READ);
 229
 230       /* XXX: what if user's constant buffer is too small?
 231        */
 232       memcpy(&buf[offset], value, nr * 4 * sizeof(float));
 233
 234       screen->buffer_unmap( screen, brw->curr.vertex_constants );
 235    }
 236
 237    if (BRW_DEBUG & DEBUG_CURBE) {
 238       for (i = 0; i < sz*16; i+=4)
 239          debug_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 240                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 241
 242       debug_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 243                    brw->curbe.last_buf, buf,
 244                    bufsz, brw->curbe.last_bufsz,
 245                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 246    }
 247
 248    if (brw->curbe.curbe_bo != NULL &&
 249        brw->curbe.last_buf &&
 250        bufsz == brw->curbe.last_bufsz &&
 251        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 252       /* constants have not changed */
 253       FREE(buf);
 254    }
 255    else {
 256       /* constants have changed */
 257       if (brw->curbe.last_buf)
 258          FREE(brw->curbe.last_buf);
 259
 260       brw->curbe.last_buf = buf;
 261       brw->curbe.last_bufsz = bufsz;
 262
 263       if (brw->curbe.curbe_bo != NULL &&
 264           (brw->curbe.need_new_bo ||
 265            brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
 266       {
 267          brw->sws->bo_unreference(brw->curbe.curbe_bo);
 268          brw->curbe.curbe_bo = NULL;
 269       }
 270
 271       if (brw->curbe.curbe_bo == NULL) {
 272          /* Allocate a single page for CURBE entries for this batchbuffer.
 273           * They're generally around 64b.
 274           */
 275          brw->curbe.curbe_bo = brw->sws->bo_alloc(brw->sws,
 276                                                   BRW_BUFFER_TYPE_CURBE,
 277                                                   4096, 1 << 6);
 278          brw->curbe.curbe_next_offset = 0;
 279       }
 280
 281       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 282       brw->curbe.curbe_next_offset += bufsz;
 283       brw->curbe.curbe_next_offset = align(brw->curbe.curbe_next_offset, 64);
 284
 285       /* Copy data to the buffer:
 286        */
 287       brw->sws->bo_subdata(brw->curbe.curbe_bo,
 288                            brw->curbe.curbe_offset,
 289                            bufsz,
 290                            buf);
 291    }
 292
 293    brw_add_validated_bo(brw, brw->curbe.curbe_bo);
 294
 295    /* Because this provokes an action (ie copy the constants into the
 296     * URB), it shouldn't be shortcircuited if identical to the
 297     * previous time - because eg. the urb destination may have
 298     * changed, or the urb contents different to last time.
 299     *
 300     * Note that the data referred to is actually copied internally,
 301     * not just used in place according to passed pointer.
 302     *
 303     * It appears that the CS unit takes care of using each available
 304     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 305     * flushes as necessary when doublebuffering of CURBEs isn't
 306     * possible.
 307     */
 308 }
 309
 310 static void emit_constant_buffer(struct brw_context *brw)
 311 {
 312    GLuint sz = brw->curbe.total_size;
 313
 314    BEGIN_BATCH(2, IGNORE_CLIPRECTS);
 315    if (sz == 0) {
 316       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 317       OUT_BATCH(0);
 318    } else {
 319       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 320       OUT_RELOC(brw->curbe.curbe_bo,
 321                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 322                 (sz - 1) + brw->curbe.curbe_offset);
 323    }
 324    ADVANCE_BATCH();
 325 }
 326
 327 const struct brw_tracked_state brw_constant_buffer = {
 328    .dirty = {
 329       .mesa = (PIPE_NEW_FRAGMENT_CONSTANTS |
 330                PIPE_NEW_VERTEX_CONSTANTS |
 331                PIPE_NEW_CLIP),
 332       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 333                BRW_NEW_VERTEX_PROGRAM |
 334                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 335                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 336                BRW_NEW_CURBE_OFFSETS |
 337                BRW_NEW_BATCH),
 338       .cache = (CACHE_NEW_WM_PROG)
 339    },
 340    .prepare = prepare_constant_buffer,
 341    .emit = emit_constant_buffer,
 342 };
 343