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_print.h"
  40 #include "shader/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    GLcontext *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          _mesa_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    GLcontext *ctx = &brw->intel.ctx;
 183    const struct brw_vertex_program *vp =
 184       brw_vertex_program_const(brw->vertex_program);
 185    const struct brw_fragment_program *fp =
 186       brw_fragment_program_const(brw->fragment_program);
 187    const GLuint sz = brw->curbe.total_size;
 188    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 189    GLfloat *buf;
 190    GLuint i;
 191
 192    if (sz == 0) {
 193       if (brw->curbe.last_buf) {
 194          free(brw->curbe.last_buf);
 195          brw->curbe.last_buf = NULL;
 196          brw->curbe.last_bufsz  = 0;
 197       }
 198       return;
 199    }
 200
 201    buf = (GLfloat *) _mesa_calloc(bufsz);
 202
 203    /* fragment shader constants */
 204    if (brw->curbe.wm_size) {
 205       GLuint offset = brw->curbe.wm_start * 16;
 206
 207       _mesa_load_state_parameters(ctx, fp->program.Base.Parameters);
 208
 209       /* copy float constants */
 210       for (i = 0; i < brw->wm.prog_data->nr_params; i++)
 211          buf[offset + i] = *brw->wm.prog_data->param[i];
 212    }
 213
 214
 215    /* The clipplanes are actually delivered to both CLIP and VS units.
 216     * VS uses them to calculate the outcode bitmasks.
 217     */
 218    if (brw->curbe.clip_size) {
 219       GLuint offset = brw->curbe.clip_start * 16;
 220       GLuint j;
 221
 222       /* If any planes are going this way, send them all this way:
 223        */
 224       for (i = 0; i < 6; i++) {
 225          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 226          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 227          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 228          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 229       }
 230
 231       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 232        * clip-space:
 233        */
 234       assert(MAX_CLIP_PLANES == 6);
 235       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 236          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 237             buf[offset + i * 4 + 0] = ctx->Transform._ClipUserPlane[j][0];
 238             buf[offset + i * 4 + 1] = ctx->Transform._ClipUserPlane[j][1];
 239             buf[offset + i * 4 + 2] = ctx->Transform._ClipUserPlane[j][2];
 240             buf[offset + i * 4 + 3] = ctx->Transform._ClipUserPlane[j][3];
 241             i++;
 242          }
 243       }
 244    }
 245
 246    /* vertex shader constants */
 247    if (brw->curbe.vs_size) {
 248       GLuint offset = brw->curbe.vs_start * 16;
 249       GLuint nr = brw->vs.prog_data->nr_params / 4;
 250
 251       if (brw->vertex_program->IsNVProgram)
 252          _mesa_load_tracked_matrices(ctx);
 253
 254       /* Updates the ParamaterValues[i] pointers for all parameters of the
 255        * basic type of PROGRAM_STATE_VAR.
 256        */
 257       _mesa_load_state_parameters(ctx, vp->program.Base.Parameters);
 258
 259       if (vp->use_const_buffer) {
 260          /* Load the subset of push constants that will get used when
 261           * we also have a pull constant buffer.
 262           */
 263          for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) {
 264             if (brw->vs.constant_map[i] != -1) {
 265                assert(brw->vs.constant_map[i] <= nr);
 266                memcpy(buf + offset + brw->vs.constant_map[i] * 4,
 267                       vp->program.Base.Parameters->ParameterValues[i],
 268                       4 * sizeof(float));
 269             }
 270          }
 271       } else {
 272          for (i = 0; i < nr; i++) {
 273             memcpy(buf + offset + i * 4,
 274                    vp->program.Base.Parameters->ParameterValues[i],
 275                    4 * sizeof(float));
 276          }
 277       }
 278    }
 279
 280    if (0) {
 281       for (i = 0; i < sz*16; i+=4)
 282          _mesa_printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 283                       buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 284
 285       _mesa_printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 286                    brw->curbe.last_buf, buf,
 287                    bufsz, brw->curbe.last_bufsz,
 288                    brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 289    }
 290
 291    if (brw->curbe.curbe_bo != NULL &&
 292        brw->curbe.last_buf &&
 293        bufsz == brw->curbe.last_bufsz &&
 294        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 295       /* constants have not changed */
 296       _mesa_free(buf);
 297    }
 298    else {
 299       /* constants have changed */
 300       if (brw->curbe.last_buf)
 301          _mesa_free(brw->curbe.last_buf);
 302
 303       brw->curbe.last_buf = buf;
 304       brw->curbe.last_bufsz = bufsz;
 305
 306       if (brw->curbe.curbe_bo != NULL &&
 307           (brw->curbe.need_new_bo ||
 308            brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size))
 309       {
 310          dri_bo_unreference(brw->curbe.curbe_bo);
 311          brw->curbe.curbe_bo = NULL;
 312       }
 313
 314       if (brw->curbe.curbe_bo == NULL) {
 315          /* Allocate a single page for CURBE entries for this batchbuffer.
 316           * They're generally around 64b.
 317           */
 318          brw->curbe.curbe_bo = dri_bo_alloc(brw->intel.bufmgr, "CURBE",
 319                                             4096, 1 << 6);
 320          brw->curbe.curbe_next_offset = 0;
 321       }
 322
 323       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 324       brw->curbe.curbe_next_offset += bufsz;
 325       brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 326
 327       /* Copy data to the buffer:
 328        */
 329       dri_bo_subdata(brw->curbe.curbe_bo, brw->curbe.curbe_offset, bufsz, buf);
 330    }
 331
 332    brw_add_validated_bo(brw, brw->curbe.curbe_bo);
 333
 334    /* Because this provokes an action (ie copy the constants into the
 335     * URB), it shouldn't be shortcircuited if identical to the
 336     * previous time - because eg. the urb destination may have
 337     * changed, or the urb contents different to last time.
 338     *
 339     * Note that the data referred to is actually copied internally,
 340     * not just used in place according to passed pointer.
 341     *
 342     * It appears that the CS unit takes care of using each available
 343     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 344     * flushes as necessary when doublebuffering of CURBEs isn't
 345     * possible.
 346     */
 347 }
 348
 349 static void emit_constant_buffer(struct brw_context *brw)
 350 {
 351    struct intel_context *intel = &brw->intel;
 352    GLuint sz = brw->curbe.total_size;
 353
 354    BEGIN_BATCH(2);
 355    if (sz == 0) {
 356       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 357       OUT_BATCH(0);
 358    } else {
 359       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 360       OUT_RELOC(brw->curbe.curbe_bo,
 361                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 362                 (sz - 1) + brw->curbe.curbe_offset);
 363    }
 364    ADVANCE_BATCH();
 365 }
 366
 367 /* This tracked state is unique in that the state it monitors varies
 368  * dynamically depending on the parameters tracked by the fragment and
 369  * vertex programs.  This is the template used as a starting point,
 370  * each context will maintain a copy of this internally and update as
 371  * required.
 372  */
 373 const struct brw_tracked_state brw_constant_buffer = {
 374    .dirty = {
 375       .mesa = _NEW_PROGRAM_CONSTANTS,
 376       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 377                BRW_NEW_VERTEX_PROGRAM |
 378                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 379                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 380                BRW_NEW_CURBE_OFFSETS |
 381                BRW_NEW_BATCH),
 382       .cache = (CACHE_NEW_WM_PROG)
 383    },
 384    .prepare = prepare_constant_buffer,
 385    .emit = emit_constant_buffer,
 386 };
 387