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 "main/glheader.h"
  35 #include "main/context.h"
  36 #include "main/macros.h"
  37 #include "main/enums.h"
  38 #include "shader/prog_parameter.h"
  39 #include "shader/prog_statevars.h"
  40 #include "intel_batchbuffer.h"
  41 #include "intel_regions.h"
  42 #include "brw_context.h"
  43 #include "brw_defines.h"
  44 #include "brw_state.h"
  45 #include "brw_util.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    GLcontext *ctx = &brw->intel.ctx;
  53    /* CACHE_NEW_WM_PROG */
  54    GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16;
  55
  56    /* BRW_NEW_VERTEX_PROGRAM */
  57    const struct brw_vertex_program *vp = brw_vertex_program_const(brw->vertex_program);
  58    GLuint nr_vp_regs = (vp->program.Base.Parameters->NumParameters * 4 + 15) / 16;
  59    GLuint nr_clip_regs = 0;
  60    GLuint total_regs;
  61
  62    /* _NEW_TRANSFORM */
  63    if (ctx->Transform.ClipPlanesEnabled) {
  64       GLuint nr_planes = 6 + brw_count_bits(ctx->Transform.ClipPlanesEnabled);
  65       nr_clip_regs = (nr_planes * 4 + 15) / 16;
  66    }
  67
  68
  69    total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs;
  70
  71    /* This can happen - what to do?  Probably rather than falling
  72     * back, the best thing to do is emit programs which code the
  73     * constants as immediate values.  Could do this either as a static
  74     * cap on WM and VS, or adaptively.
  75     *
  76     * Unfortunately, this is currently dependent on the results of the
  77     * program generation process (in the case of wm), so this would
  78     * introduce the need to re-generate programs in the event of a
  79     * curbe allocation failure.
  80     */
  81    /* Max size is 32 - just large enough to
  82     * hold the 128 parameters allowed by
  83     * the fragment and vertex program
  84     * api's.  It's not clear what happens
  85     * when both VP and FP want to use 128
  86     * parameters, though.
  87     */
  88    assert(total_regs <= 32);
  89
  90    /* Lazy resize:
  91     */
  92    if (nr_fp_regs > brw->curbe.wm_size ||
  93        nr_vp_regs > brw->curbe.vs_size ||
  94        nr_clip_regs != brw->curbe.clip_size ||
  95        (total_regs < brw->curbe.total_size / 4 &&
  96         brw->curbe.total_size > 16)) {
  97
  98       GLuint reg = 0;
  99
 100       /* Calculate a new layout:
 101        */
 102       reg = 0;
 103       brw->curbe.wm_start = reg;
 104       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
 105       brw->curbe.clip_start = reg;
 106       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
 107       brw->curbe.vs_start = reg;
 108       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
 109       brw->curbe.total_size = reg;
 110
 111       if (0)
 112          _mesa_printf("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
 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       .mesa = _NEW_TRANSFORM,
 128       .brw  = BRW_NEW_VERTEX_PROGRAM,
 129       .cache = CACHE_NEW_WM_PROG
 130    },
 131    .prepare = calculate_curbe_offsets
 132 };
 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_cs_urb_state(struct brw_context *brw)
 143 {
 144    struct brw_cs_urb_state cs_urb;
 145    memset(&cs_urb, 0, sizeof(cs_urb));
 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    cs_urb.header.opcode = CMD_CS_URB_STATE;
 152    cs_urb.header.length = sizeof(cs_urb)/4 - 2;
 153
 154    /* BRW_NEW_URB_FENCE */
 155    cs_urb.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
 156    cs_urb.bits0.urb_entry_size = brw->urb.csize - 1;
 157
 158    assert(brw->urb.nr_cs_entries);
 159    BRW_CACHED_BATCH_STRUCT(brw, &cs_urb);
 160 }
 161
 162 static GLfloat 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 prepare_constant_buffer(struct brw_context *brw)
 176 {
 177    GLcontext *ctx = &brw->intel.ctx;
 178    const struct brw_vertex_program *vp =
 179       brw_vertex_program_const(brw->vertex_program);
 180    const struct brw_fragment_program *fp =
 181       brw_fragment_program_const(brw->fragment_program);
 182    const GLuint sz = brw->curbe.total_size;
 183    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 184    GLfloat *buf;
 185    GLuint i;
 186
 187    /* Update our own dependency flags.  This works because this
 188     * function will also be called whenever fp or vp changes.
 189     */
 190    brw->curbe.tracked_state.dirty.mesa = (_NEW_TRANSFORM|_NEW_PROJECTION);
 191    brw->curbe.tracked_state.dirty.mesa |= vp->program.Base.Parameters->StateFlags;
 192    brw->curbe.tracked_state.dirty.mesa |= fp->program.Base.Parameters->StateFlags;
 193
 194    if (sz == 0) {
 195       if (brw->curbe.last_buf) {
 196          free(brw->curbe.last_buf);
 197          brw->curbe.last_buf = NULL;
 198          brw->curbe.last_bufsz  = 0;
 199       }
 200       return;
 201    }
 202
 203    buf = (GLfloat *) _mesa_calloc(bufsz);
 204
 205    /* fragment shader constants */
 206    if (brw->curbe.wm_size) {
 207       GLuint offset = brw->curbe.wm_start * 16;
 208
 209       _mesa_load_state_parameters(ctx, fp->program.Base.Parameters);
 210
 211       /* copy float constants */
 212       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 213          buf[offset + i] = *brw->wm.prog_data->param[i];
 214    }
 215
 216
 217    /* The clipplanes are actually delivered to both CLIP and VS units.
 218     * VS uses them to calculate the outcode bitmasks.
 219     */
 220    if (brw->curbe.clip_size) {
 221       GLuint offset = brw->curbe.clip_start * 16;
 222       GLuint j;
 223
 224       /* If any planes are going this way, send them all this way:
 225        */
 226       for (i = 0; i < 6; i++) {
 227          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 228          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 229          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 230          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 231       }
 232
 233       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 234        * clip-space:
 235        */
 236       assert(MAX_CLIP_PLANES == 6);
 237       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 238          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 239             buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0];
 240             buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1];
 241             buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2];
 242             buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3];
 243             i++;
 244          }
 245       }
 246    }
 247
 248    /* vertex shader constants */
 249    if (brw->curbe.vs_size) {
 250       GLuint offset = brw->curbe.vs_start * 16;
 251       GLuint nr = vp->program.Base.Parameters->NumParameters;
 252
 253       _mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
 254
 255       /* XXX just use a memcpy here */
 256       for (i = 0; i < nr; i++) {
 257          const GLfloat *value = vp->program.Base.Parameters->ParameterValues[i];
 258          buf[offset + i * 4 + 0] = value[0];
 259          buf[offset + i * 4 + 1] = value[1];
 260          buf[offset + i * 4 + 2] = value[2];
 261          buf[offset + i * 4 + 3] = value[3];
 262       }
 263    }
 264
 265    if (0) {
 266       for (i = 0; i < sz*16; i+=4)
 267          _mesa_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 268                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 269
 270       _mesa_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 271                    brw->curbe.last_buf, buf,
 272                    bufsz, brw->curbe.last_bufsz,
 273                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 274    }
 275
 276    if (brw->curbe.curbe_bo != NULL &&
 277        brw->curbe.last_buf &&
 278        bufsz == brw->curbe.last_bufsz &&
 279        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 280       /* constants have not changed */
 281       _mesa_free(buf);
 282    }
 283    else {
 284       /* constants have changed */
 285       if (brw->curbe.last_buf)
 286          _mesa_free(brw->curbe.last_buf);
 287
 288       brw->curbe.last_buf = buf;
 289       brw->curbe.last_bufsz = bufsz;
 290
 291       if (brw->curbe.curbe_bo != NULL &&
 292           (brw->curbe.need_new_bo ||
 293            brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
 294       {
 295          dri_bo_unreference(brw->curbe.curbe_bo);
 296          brw->curbe.curbe_bo = NULL;
 297       }
 298
 299       if (brw->curbe.curbe_bo == NULL) {
 300          /* Allocate a single page for CURBE entries for this batchbuffer.
 301           * They're generally around 64b.
 302           */
 303          brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
 304                                             4096, 1 << 6);
 305          brw->curbe.curbe_next_offset = 0;
 306       }
 307
 308       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 309       brw->curbe.curbe_next_offset += bufsz;
 310       brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 311
 312       /* Copy data to the buffer:
 313        */
 314       dri_bo_subdata(brw->curbe.curbe_bo, brw->curbe.curbe_offset, bufsz, buf);
 315    }
 316
 317    brw_add_validated_bo(brw, brw->curbe.curbe_bo);
 318
 319    /* Because this provokes an action (ie copy the constants into the
 320     * URB), it shouldn't be shortcircuited if identical to the
 321     * previous time - because eg. the urb destination may have
 322     * changed, or the urb contents different to last time.
 323     *
 324     * Note that the data referred to is actually copied internally,
 325     * not just used in place according to passed pointer.
 326     *
 327     * It appears that the CS unit takes care of using each available
 328     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 329     * flushes as necessary when doublebuffering of CURBEs isn't
 330     * possible.
 331     */
 332 }
 333
 334
 335 /**
 336  * Vertex/fragment shader constants are stored in a pseudo 1D texture.
 337  * This function updates the constants in that buffer.
 338  */
 339 static void
 340 update_texture_constant_buffer(struct brw_context *brw)
 341 {
 342    struct brw_fragment_program *fp =
 343       (struct brw_fragment_program *) brw->fragment_program;
 344    const struct gl_program_parameter_list *params = fp->program.Base.Parameters;
 345    const int size = params->NumParameters * 4 * sizeof(GLfloat);
 346
 347    assert(fp->const_buffer);
 348    assert(fp->const_buffer->size >= size);
 349
 350    /* copy constants into the buffer */
 351    if (size > 0) {
 352       GLubyte *map;
 353       dri_bo_map(fp->const_buffer, GL_TRUE);
 354       map = fp->const_buffer->virtual;
 355       memcpy(map, params->ParameterValues, size);
 356       dri_bo_unmap(fp->const_buffer);
 357    }
 358 }
 359
 360
 361 static void emit_constant_buffer(struct brw_context *brw)
 362 {
 363    struct intel_context *intel = &brw->intel;
 364    GLuint sz = brw->curbe.total_size;
 365
 366    update_texture_constant_buffer(brw);
 367
 368    BEGIN_BATCH(2, IGNORE_CLIPRECTS);
 369    if (sz == 0) {
 370       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 371       OUT_BATCH(0);
 372    } else {
 373       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 374       OUT_RELOC(brw->curbe.curbe_bo,
 375                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 376                 (sz - 1) + brw->curbe.curbe_offset);
 377    }
 378    ADVANCE_BATCH();
 379 }
 380
 381 /* This tracked state is unique in that the state it monitors varies
 382  * dynamically depending on the parameters tracked by the fragment and
 383  * vertex programs.  This is the template used as a starting point,
 384  * each context will maintain a copy of this internally and update as
 385  * required.
 386  */
 387 const struct brw_tracked_state brw_constant_buffer = {
 388    .dirty = {
 389       .mesa = (_NEW_TRANSFORM|_NEW_PROJECTION),      /* plus fp and vp flags */
 390       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 391                BRW_NEW_VERTEX_PROGRAM |
 392                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 393                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 394                BRW_NEW_CURBE_OFFSETS |
 395                BRW_NEW_BATCH),
 396       .cache = (CACHE_NEW_WM_PROG)
 397    },
 398    .prepare = prepare_constant_buffer,
 399    .emit = emit_constant_buffer,
 400 };
 401