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 "glheader.h"
  35 #include "context.h"
  36 #include "macros.h"
  37 #include "enums.h"
  38 #include "shader/program.h"
  39 #include "intel_batchbuffer.h"
  40 #include "brw_context.h"
  41 #include "brw_defines.h"
  42 #include "brw_state.h"
  43 #include "brw_util.h"
  44 #include "brw_aub.h"
  45
  46
  47 /* Partition the CURBE between the various users of constant values:
  48  */
  49 static void calculate_curbe_offsets( struct brw_context *brw )
  50 {
  51    /* CACHE_NEW_WM_PROG */
  52    GLuint nr_fp_regs = (brw->wm.prog_data->nr_params + 15) / 16;
  53
  54    /* BRW_NEW_VERTEX_PROGRAM */
  55    struct brw_vertex_program *vp = (struct brw_vertex_program *)brw->vertex_program;
  56    GLuint nr_vp_regs = (vp->program.Base.Parameters->NumParameters * 4 + 15) / 16;
  57    GLuint nr_clip_regs = 0;
  58    GLuint total_regs;
  59
  60    /* _NEW_TRANSFORM */
  61    if (brw->attribs.Transform->ClipPlanesEnabled) {
  62       GLuint nr_planes = 6 + brw_count_bits(brw->attribs.Transform->ClipPlanesEnabled);
  63       nr_clip_regs = (nr_planes * 4 + 15) / 16;
  64    }
  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       GLuint 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          _mesa_printf("curbe wm %d+%d clip %d+%d vs %d+%d\n",
 111                       brw->curbe.wm_start,
 112                       brw->curbe.wm_size,
 113                       brw->curbe.clip_start,
 114                       brw->curbe.clip_size,
 115                       brw->curbe.vs_start,
 116                       brw->curbe.vs_size );
 117
 118       brw->state.dirty.brw |= BRW_NEW_CURBE_OFFSETS;
 119    }
 120 }
 121
 122
 123 const struct brw_tracked_state brw_curbe_offsets = {
 124    .dirty = {
 125       .mesa = _NEW_TRANSFORM,
 126       .brw  = BRW_NEW_VERTEX_PROGRAM,
 127       .cache = CACHE_NEW_WM_PROG
 128    },
 129    .update = calculate_curbe_offsets
 130 };
 131
 132
 133
 134
 135 /* Define the number of curbes within CS's urb allocation.  Multiple
 136  * urb entries -> multiple curbes.  These will be used by
 137  * fixed-function hardware in a double-buffering scheme to avoid a
 138  * pipeline stall each time the contents of the curbe is changed.
 139  */
 140 void brw_upload_constant_buffer_state(struct brw_context *brw)
 141 {
 142    struct brw_constant_buffer_state cbs;
 143    memset(&cbs, 0, sizeof(cbs));
 144
 145    /* It appears that this is the state packet for the CS unit, ie. the
 146     * urb entries detailed here are housed in the CS range from the
 147     * URB_FENCE command.
 148     */
 149    cbs.header.opcode = CMD_CONST_BUFFER_STATE;
 150    cbs.header.length = sizeof(cbs)/4 - 2;
 151
 152    /* BRW_NEW_URB_FENCE */
 153    cbs.bits0.nr_urb_entries = brw->urb.nr_cs_entries;
 154    cbs.bits0.urb_entry_size = brw->urb.csize - 1;
 155
 156    assert(brw->urb.nr_cs_entries);
 157    BRW_CACHED_BATCH_STRUCT(brw, &cbs);
 158 }
 159
 160 #if 0
 161 const struct brw_tracked_state brw_constant_buffer_state = {
 162    .dirty = {
 163       .mesa = 0,
 164       .brw = BRW_NEW_URB_FENCE,
 165       .cache = 0
 166    },
 167    .update = brw_upload_constant_buffer_state
 168 };
 169 #endif
 170
 171
 172 static GLfloat fixed_plane[6][4] = {
 173    { 0,    0,   -1, 1 },
 174    { 0,    0,    1, 1 },
 175    { 0,   -1,    0, 1 },
 176    { 0,    1,    0, 1 },
 177    {-1,    0,    0, 1 },
 178    { 1,    0,    0, 1 }
 179 };
 180
 181 /* Upload a new set of constants.  Too much variability to go into the
 182  * cache mechanism, but maybe would benefit from a comparison against
 183  * the current uploaded set of constants.
 184  */
 185 static void upload_constant_buffer(struct brw_context *brw)
 186 {
 187    GLcontext *ctx = &brw->intel.ctx;
 188    struct brw_vertex_program *vp = (struct brw_vertex_program *)brw->vertex_program;
 189    struct brw_fragment_program *fp = (struct brw_fragment_program *)brw->fragment_program;
 190    struct brw_mem_pool *pool = &brw->pool[BRW_GS_POOL];
 191    GLuint sz = brw->curbe.total_size;
 192    GLuint bufsz = sz * 16 * sizeof(GLfloat);
 193    GLfloat *buf;
 194    GLuint i;
 195
 196    /* Update our own dependency flags.  This works because this
 197     * function will also be called whenever fp or vp changes.
 198     */
 199    brw->curbe.tracked_state.dirty.mesa = (_NEW_TRANSFORM|_NEW_PROJECTION);
 200    brw->curbe.tracked_state.dirty.mesa |= vp->param_state;
 201    brw->curbe.tracked_state.dirty.mesa |= fp->param_state;
 202
 203    if (sz == 0) {
 204       struct brw_constant_buffer cb;
 205       cb.header.opcode = CMD_CONST_BUFFER;
 206       cb.header.length = sizeof(cb)/4 - 2;
 207       cb.header.valid = 0;
 208       cb.bits0.buffer_length = 0;
 209       cb.bits0.buffer_address = 0;
 210       BRW_BATCH_STRUCT(brw, &cb);
 211
 212       if (brw->curbe.last_buf) {
 213          free(brw->curbe.last_buf);
 214          brw->curbe.last_buf = NULL;
 215          brw->curbe.last_bufsz  = 0;
 216       }
 217
 218       return;
 219    }
 220
 221    buf = (GLfloat *)malloc(bufsz);
 222
 223    memset(buf, 0, bufsz);
 224
 225    if (brw->curbe.wm_size) {
 226       GLuint offset = brw->curbe.wm_start * 16;
 227
 228       _mesa_load_state_parameters(ctx, fp->program.Base.Parameters);
 229
 230       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 231          buf[offset + i] = brw->wm.prog_data->param[i][0];
 232    }
 233
 234
 235    /* The clipplanes are actually delivered to both CLIP and VS units.
 236     * VS uses them to calculate the outcode bitmasks.
 237     */
 238    if (brw->curbe.clip_size) {
 239       GLuint offset = brw->curbe.clip_start * 16;
 240       GLuint j;
 241
 242       /* If any planes are going this way, send them all this way:
 243        */
 244       for (i = 0; i < 6; i++) {
 245          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 246          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 247          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 248          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 249       }
 250
 251       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 252        * clip-space:
 253        */
 254       assert(MAX_CLIP_PLANES == 6);
 255       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 256          if (brw->attribs.Transform->ClipPlanesEnabled & (1<<j)) {
 257             buf[offset + i * 4 + 0] = brw->attribs.Transform->_ClipUserPlane[j][0];
 258             buf[offset + i * 4 + 1] = brw->attribs.Transform->_ClipUserPlane[j][1];
 259             buf[offset + i * 4 + 2] = brw->attribs.Transform->_ClipUserPlane[j][2];
 260             buf[offset + i * 4 + 3] = brw->attribs.Transform->_ClipUserPlane[j][3];
 261             i++;
 262          }
 263       }
 264    }
 265
 266
 267    if (brw->curbe.vs_size) {
 268       GLuint offset = brw->curbe.vs_start * 16;
 269       GLuint nr = vp->program.Base.Parameters->NumParameters;
 270
 271       _mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
 272
 273       for (i = 0; i < nr; i++) {
 274          buf[offset + i * 4 + 0] = vp->program.Base.Parameters->ParameterValues[i][0];
 275          buf[offset + i * 4 + 1] = vp->program.Base.Parameters->ParameterValues[i][1];
 276          buf[offset + i * 4 + 2] = vp->program.Base.Parameters->ParameterValues[i][2];
 277          buf[offset + i * 4 + 3] = vp->program.Base.Parameters->ParameterValues[i][3];
 278       }
 279    }
 280
 281    if (0) {
 282       for (i = 0; i < sz*16; i+=4)
 283          _mesa_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 284                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 285
 286       _mesa_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 287                    brw->curbe.last_buf, buf,
 288                    bufsz, brw->curbe.last_bufsz,
 289                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 290    }
 291
 292    if (brw->curbe.last_buf &&
 293        bufsz == brw->curbe.last_bufsz &&
 294        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 295       free(buf);
 296 /*       return; */
 297    }
 298    else {
 299       if (brw->curbe.last_buf)
 300          free(brw->curbe.last_buf);
 301       brw->curbe.last_buf = buf;
 302       brw->curbe.last_bufsz = bufsz;
 303
 304
 305       if (!brw_pool_alloc(pool,
 306                           bufsz,
 307                           6,
 308                           &brw->curbe.gs_offset)) {
 309          _mesa_printf("out of GS memory for curbe\n");
 310          assert(0);
 311          return;
 312       }
 313
 314
 315       /* Copy data to the buffer:
 316        */
 317       bmBufferSubDataAUB(&brw->intel,
 318                          pool->buffer,
 319                          brw->curbe.gs_offset,
 320                          bufsz,
 321                          buf,
 322                          DW_CONSTANT_BUFFER,
 323                          0);
 324    }
 325
 326    /* TODO: only emit the constant_buffer packet when necessary, ie:
 327       - contents have changed
 328       - offset has changed
 329       - hw requirements due to other packets emitted.
 330    */
 331    {
 332       struct brw_constant_buffer cb;
 333
 334       memset(&cb, 0, sizeof(cb));
 335
 336       cb.header.opcode = CMD_CONST_BUFFER;
 337       cb.header.length = sizeof(cb)/4 - 2;
 338       cb.header.valid = 1;
 339       cb.bits0.buffer_length = sz - 1;
 340       cb.bits0.buffer_address = brw->curbe.gs_offset >> 6;
 341
 342       /* Because this provokes an action (ie copy the constants into the
 343        * URB), it shouldn't be shortcircuited if identical to the
 344        * previous time - because eg. the urb destination may have
 345        * changed, or the urb contents different to last time.
 346        *
 347        * Note that the data referred to is actually copied internally,
 348        * not just used in place according to passed pointer.
 349        *
 350        * It appears that the CS unit takes care of using each available
 351        * URB entry (Const URB Entry == CURBE) in turn, and issuing
 352        * flushes as necessary when doublebuffering of CURBEs isn't
 353        * possible.
 354        */
 355 /*       intel_batchbuffer_align(brw->intel.batch, 64, sizeof(cb)); */
 356       BRW_BATCH_STRUCT(brw, &cb);
 357 /*       intel_batchbuffer_align(brw->intel.batch, 64, 0); */
 358    }
 359 }
 360
 361 /* This tracked state is unique in that the state it monitors varies
 362  * dynamically depending on the parameters tracked by the fragment and
 363  * vertex programs.  This is the template used as a starting point,
 364  * each context will maintain a copy of this internally and update as
 365  * required.
 366  */
 367 const struct brw_tracked_state brw_constant_buffer = {
 368    .dirty = {
 369       .mesa = (_NEW_TRANSFORM|_NEW_PROJECTION),      /* plus fp and vp flags */
 370       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 371                BRW_NEW_VERTEX_PROGRAM |
 372                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 373                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 374                BRW_NEW_CURBE_OFFSETS),
 375       .cache = (CACHE_NEW_WM_PROG)
 376    },
 377    .update = upload_constant_buffer
 378 };
 379