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 "program/prog_parameter.h"
  39 #include "program/prog_print.h"
  40 #include "program/prog_statevars.h"
  41 #include "intel_batchbuffer.h"
  42 #include "intel_regions.h"
  43 #include "brw_context.h"
  44 #include "brw_defines.h"
  45 #include "brw_state.h"
  46 #include "brw_util.h"
  47
  48
  49 /**
  50  * Partition the CURBE between the various users of constant values:
  51  * Note that vertex and fragment shaders can now fetch constants out
  52  * of constant buffers.  We no longer allocatea block of the GRF for
  53  * constants.  That greatly reduces the demand for space in the CURBE.
  54  * Some of the comments within are dated...
  55  */
  56 static void calculate_curbe_offsets( struct brw_context *brw )
  57 {
  58    struct gl_context *ctx = &brw->intel.ctx;
  59    /* CACHE_NEW_WM_PROG */
  60    const GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16;
  61
  62    /* BRW_NEW_VERTEX_PROGRAM */
  63    const GLuint nr_vp_regs = (brw->vs.prog_data->nr_params + 15) / 16;
  64    GLuint nr_clip_regs = 0;
  65    GLuint total_regs;
  66
  67    /* _NEW_TRANSFORM */
  68    if (ctx->Transform.ClipPlanesEnabled) {
  69       GLuint nr_planes = 6 + brw_count_bits(ctx->Transform.ClipPlanesEnabled);
  70       nr_clip_regs = (nr_planes * 4 + 15) / 16;
  71    }
  72
  73
  74    total_regs = nr_fp_regs + nr_vp_regs + nr_clip_regs;
  75
  76    /* This can happen - what to do?  Probably rather than falling
  77     * back, the best thing to do is emit programs which code the
  78     * constants as immediate values.  Could do this either as a static
  79     * cap on WM and VS, or adaptively.
  80     *
  81     * Unfortunately, this is currently dependent on the results of the
  82     * program generation process (in the case of wm), so this would
  83     * introduce the need to re-generate programs in the event of a
  84     * curbe allocation failure.
  85     */
  86    /* Max size is 32 - just large enough to
  87     * hold the 128 parameters allowed by
  88     * the fragment and vertex program
  89     * api's.  It's not clear what happens
  90     * when both VP and FP want to use 128
  91     * parameters, though.
  92     */
  93    assert(total_regs <= 32);
  94
  95    /* Lazy resize:
  96     */
  97    if (nr_fp_regs > brw->curbe.wm_size ||
  98        nr_vp_regs > brw->curbe.vs_size ||
  99        nr_clip_regs != brw->curbe.clip_size ||
 100        (total_regs < brw->curbe.total_size / 4 &&
 101         brw->curbe.total_size > 16)) {
 102
 103       GLuint reg = 0;
 104
 105       /* Calculate a new layout:
 106        */
 107       reg = 0;
 108       brw->curbe.wm_start = reg;
 109       brw->curbe.wm_size = nr_fp_regs; reg += nr_fp_regs;
 110       brw->curbe.clip_start = reg;
 111       brw->curbe.clip_size = nr_clip_regs; reg += nr_clip_regs;
 112       brw->curbe.vs_start = reg;
 113       brw->curbe.vs_size = nr_vp_regs; reg += nr_vp_regs;
 114       brw->curbe.total_size = reg;
 115
 116       if (0)
 117          printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
 118                 brw->curbe.wm_start,
 119                 brw->curbe.wm_size,
 120                 brw->curbe.clip_start,
 121                 brw->curbe.clip_size,
 122                 brw->curbe.vs_start,
 123                 brw->curbe.vs_size );
 124
 125       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 126    }
 127 }
 128
 129
 130 const struct brw_tracked_state brw_curbe_offsets = {
 131    .dirty = {
 132       .mesa = _NEW_TRANSFORM,
 133       .brw  = BRW_NEW_VERTEX_PROGRAM | BRW_NEW_CONTEXT,
 134       .cache = CACHE_NEW_WM_PROG
 135    },
 136    .prepare = calculate_curbe_offsets
 137 };
 138
 139
 140
 141
 142 /* Define the number of curbes within CS's urb allocation.  Multiple
 143  * urb entries -> multiple curbes.  These will be used by
 144  * fixed-function hardware in a double-buffering scheme to avoid a
 145  * pipeline stall each time the contents of the curbe is changed.
 146  */
 147 void brw_upload_cs_urb_state(struct brw_context *brw)
 148 {
 149    struct brw_cs_urb_state cs_urb;
 150    memset(&cs_urb, 0, sizeof(cs_urb));
 151
 152    /* It appears that this is the state packet for the CS unit, ie. the
 153     * urb entries detailed here are housed in the CS range from the
 154     * URB_FENCE command.
 155     */
 156    cs_urb.header.opcode = CMD_CS_URB_STATE;
 157    cs_urb.header.length = sizeof(cs_urb)/4 - 2;
 158
 159    /* BRW_NEW_URB_FENCE */
 160    cs_urb.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
 161    cs_urb.bits0.urb_entry_size = brw->urb.csize - 1;
 162
 163    assert(brw->urb.nr_cs_entries);
 164    BRW_CACHED_BATCH_STRUCT(brw, &cs_urb);
 165 }
 166
 167 static GLfloat fixed_plane[6][4] = {
 168    { 0,    0,   -1, 1 },
 169    { 0,    0,    1, 1 },
 170    { 0,   -1,    0, 1 },
 171    { 0,    1,    0, 1 },
 172    {-1,    0,    0, 1 },
 173    { 1,    0,    0, 1 }
 174 };
 175
 176 /* Upload a new set of constants.  Too much variability to go into the
 177  * cache mechanism, but maybe would benefit from a comparison against
 178  * the current uploaded set of constants.
 179  */
 180 static void prepare_constant_buffer(struct brw_context *brw)
 181 {
 182    struct gl_context *ctx = &brw->intel.ctx;
 183    const struct brw_vertex_program *vp =
 184       brw_vertex_program_const(brw->vertex_program);
 185    const GLuint sz = brw->curbe.total_size;
 186    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 187    GLfloat *buf;
 188    GLuint i;
 189
 190    if (sz == 0) {
 191       brw->curbe.last_bufsz  = 0;
 192       return;
 193    }
 194
 195    buf = brw->curbe.next_buf;
 196
 197    /* fragment shader constants */
 198    if (brw->curbe.wm_size) {
 199       GLuint offset = brw->curbe.wm_start * 16;
 200
 201       /* copy float constants */
 202       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 203          buf[offset + i] = *brw->wm.prog_data->param[i];
 204    }
 205
 206
 207    /* The clipplanes are actually delivered to both CLIP and VS units.
 208     * VS uses them to calculate the outcode bitmasks.
 209     */
 210    if (brw->curbe.clip_size) {
 211       GLuint offset = brw->curbe.clip_start * 16;
 212       GLuint j;
 213
 214       /* If any planes are going this way, send them all this way:
 215        */
 216       for (i = 0; i < 6; i++) {
 217          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 218          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 219          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 220          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 221       }
 222
 223       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 224        * clip-space:
 225        */
 226       assert(MAX_CLIP_PLANES == 6);
 227       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 228          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 229             buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0];
 230             buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1];
 231             buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2];
 232             buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3];
 233             i++;
 234          }
 235       }
 236    }
 237
 238    /* vertex shader constants */
 239    if (brw->curbe.vs_size) {
 240       GLuint offset = brw->curbe.vs_start * 16;
 241       GLuint nr = brw->vs.prog_data->nr_params / 4;
 242
 243       if (vp->use_const_buffer) {
 244          /* Load the subset of push constants that will get used when
 245           * we also have a pull constant buffer.
 246           */
 247          for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) {
 248             if (brw->vs.constant_map[i] != -1) {
 249                assert(brw->vs.constant_map[i] <= nr);
 250                memcpy(buf + offset + brw->vs.constant_map[i] * 4,
 251                       vp->program.Base.Parameters->ParameterValues[i],
 252                       4 * sizeof(float));
 253             }
 254          }
 255       } else {
 256          for (i = 0; i < nr; i++) {
 257             memcpy(buf + offset + i * 4,
 258                    vp->program.Base.Parameters->ParameterValues[i],
 259                    4 * sizeof(float));
 260          }
 261       }
 262    }
 263
 264    if (0) {
 265       for (i = 0; i < sz*16; i+=4)
 266          printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 267                 buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 268
 269       printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 270              brw->curbe.last_buf, buf,
 271              bufsz, brw->curbe.last_bufsz,
 272              brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 273    }
 274
 275    if (brw->curbe.curbe_bo != NULL &&
 276        bufsz == brw->curbe.last_bufsz &&
 277        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 278       /* constants have not changed */
 279    } else {
 280       /* Update the record of what our last set of constants was.  We
 281        * don't just flip the pointers because we don't fill in the
 282        * data in the padding between the entries.
 283        */
 284       memcpy(brw->curbe.last_buf, buf, bufsz);
 285       brw->curbe.last_bufsz = bufsz;
 286
 287       if (brw->curbe.curbe_bo != NULL &&
 288           brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
 289       {
 290          drm_intel_gem_bo_unmap_gtt(brw->curbe.curbe_bo);
 291          drm_intel_bo_unreference(brw->curbe.curbe_bo);
 292          brw->curbe.curbe_bo = NULL;
 293       }
 294
 295       if (brw->curbe.curbe_bo == NULL) {
 296          /* Allocate a single page for CURBE entries for this batchbuffer.
 297           * They're generally around 64b.
 298           */
 299          brw->curbe.curbe_bo = drm_intel_bo_alloc(brw->intel.bufmgr, "CURBE",
 300                                                   4096, 1 << 6);
 301          brw->curbe.curbe_next_offset = 0;
 302          drm_intel_gem_bo_map_gtt(brw->curbe.curbe_bo);
 303          assert(bufsz < 4096);
 304       }
 305
 306       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 307       brw->curbe.curbe_next_offset += bufsz;
 308       brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 309
 310       /* Copy data to the buffer:
 311        */
 312       memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset,
 313              buf,
 314              bufsz);
 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 static void emit_constant_buffer(struct brw_context *brw)
 335 {
 336    struct intel_context *intel = &brw->intel;
 337    GLuint sz = brw->curbe.total_size;
 338
 339    BEGIN_BATCH(2);
 340    if (sz == 0) {
 341       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 342       OUT_BATCH(0);
 343    } else {
 344       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 345       OUT_RELOC(brw->curbe.curbe_bo,
 346                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 347                 (sz - 1) + brw->curbe.curbe_offset);
 348    }
 349    ADVANCE_BATCH();
 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       .mesa = _NEW_PROGRAM_CONSTANTS,
 361       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 362                BRW_NEW_VERTEX_PROGRAM |
 363                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 364                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 365                BRW_NEW_CURBE_OFFSETS |
 366                BRW_NEW_BATCH),
 367       .cache = (CACHE_NEW_WM_PROG)
 368    },
 369    .prepare = prepare_constant_buffer,
 370    .emit = emit_constant_buffer,
 371 };
 372