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 + _mesa_bitcount_64(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    .emit = 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 intel_context *intel = &brw->intel;
 150
 151    BEGIN_BATCH(2);
 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    OUT_BATCH(CMD_CS_URB_STATE << 16 | (2-2));
 157
 158    /* BRW_NEW_URB_FENCE */
 159    if (brw->urb.csize == 0) {
 160       OUT_BATCH(0);
 161    } else {
 162       /* BRW_NEW_URB_FENCE */
 163       assert(brw->urb.nr_cs_entries);
 164       OUT_BATCH((brw->urb.csize - 1) << 4 | brw->urb.nr_cs_entries);
 165    }
 166    CACHED_BATCH();
 167 }
 168
 169 static GLfloat fixed_plane[6][4] = {
 170    { 0,    0,   -1, 1 },
 171    { 0,    0,    1, 1 },
 172    { 0,   -1,    0, 1 },
 173    { 0,    1,    0, 1 },
 174    {-1,    0,    0, 1 },
 175    { 1,    0,    0, 1 }
 176 };
 177
 178 /* Upload a new set of constants.  Too much variability to go into the
 179  * cache mechanism, but maybe would benefit from a comparison against
 180  * the current uploaded set of constants.
 181  */
 182 static void
 183 brw_upload_constant_buffer(struct brw_context *brw)
 184 {
 185    struct intel_context *intel = &brw->intel;
 186    struct gl_context *ctx = &intel->ctx;
 187    const struct brw_vertex_program *vp =
 188       brw_vertex_program_const(brw->vertex_program);
 189    const GLuint sz = brw->curbe.total_size;
 190    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 191    GLfloat *buf;
 192    GLuint i;
 193    gl_clip_plane *clip_planes;
 194
 195    if (sz == 0) {
 196       brw->curbe.last_bufsz  = 0;
 197       goto emit;
 198    }
 199
 200    buf = brw->curbe.next_buf;
 201
 202    /* fragment shader constants */
 203    if (brw->curbe.wm_size) {
 204       GLuint offset = brw->curbe.wm_start * 16;
 205
 206       /* copy float constants */
 207       for (i = 0; i < brw->wm.prog_data->nr_params; i++) {
 208          buf[offset + i] = *brw->wm.prog_data->param[i];
 209       }
 210    }
 211
 212
 213    /* When using the old VS backend, the clipplanes are actually delivered to
 214     * both CLIP and VS units.  VS uses them to calculate the outcode bitmasks.
 215     *
 216     * When using the new VS backend, it is responsible for setting up its own
 217     * clipplane constants if it needs them.  This results in a slight waste of
 218     * of curbe space, but the advantage is that the new VS backend can use its
 219     * general-purpose uniform layout code to store the clipplanes.
 220     */
 221    if (brw->curbe.clip_size) {
 222       GLuint offset = brw->curbe.clip_start * 16;
 223       GLuint j;
 224
 225       /* If any planes are going this way, send them all this way:
 226        */
 227       for (i = 0; i < 6; i++) {
 228          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 229          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 230          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 231          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 232       }
 233
 234       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 235        * clip-space:
 236        */
 237       clip_planes = brw_select_clip_planes(ctx);
 238       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 239          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 240             buf[offset + i * 4 + 0] = clip_planes[j][0];
 241             buf[offset + i * 4 + 1] = clip_planes[j][1];
 242             buf[offset + i * 4 + 2] = clip_planes[j][2];
 243             buf[offset + i * 4 + 3] = clip_planes[j][3];
 244             i++;
 245          }
 246       }
 247    }
 248
 249    /* vertex shader constants */
 250    if (brw->curbe.vs_size) {
 251       GLuint offset = brw->curbe.vs_start * 16;
 252       GLuint nr = brw->vs.prog_data->nr_params / 4;
 253
 254       if (brw->vs.prog_data->uses_new_param_layout) {
 255          for (i = 0; i < brw->vs.prog_data->nr_params; i++) {
 256             buf[offset + i] = *brw->vs.prog_data->param[i];
 257          }
 258       } else {
 259          /* Load the subset of push constants that will get used when
 260           * we also have a pull constant buffer.
 261           */
 262          for (i = 0; i < vp->program.Base.Parameters->NumParameters; i++) {
 263             if (brw->vs.constant_map[i] != -1) {
 264                assert(brw->vs.constant_map[i] <= nr);
 265                memcpy(buf + offset + brw->vs.constant_map[i] * 4,
 266                       vp->program.Base.Parameters->ParameterValues[i],
 267                       4 * sizeof(float));
 268             }
 269          }
 270       }
 271    }
 272
 273    if (0) {
 274       for (i = 0; i < sz*16; i+=4)
 275          printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 276                 buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 277
 278       printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 279              brw->curbe.last_buf, buf,
 280              bufsz, brw->curbe.last_bufsz,
 281              brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 282    }
 283
 284    if (brw->curbe.curbe_bo != NULL &&
 285        bufsz == brw->curbe.last_bufsz &&
 286        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 287       /* constants have not changed */
 288    } else {
 289       /* Update the record of what our last set of constants was.  We
 290        * don't just flip the pointers because we don't fill in the
 291        * data in the padding between the entries.
 292        */
 293       memcpy(brw->curbe.last_buf, buf, bufsz);
 294       brw->curbe.last_bufsz = bufsz;
 295
 296       if (brw->curbe.curbe_bo != NULL &&
 297           brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
 298       {
 299          drm_intel_gem_bo_unmap_gtt(brw->curbe.curbe_bo);
 300          drm_intel_bo_unreference(brw->curbe.curbe_bo);
 301          brw->curbe.curbe_bo = NULL;
 302       }
 303
 304       if (brw->curbe.curbe_bo == NULL) {
 305          /* Allocate a single page for CURBE entries for this batchbuffer.
 306           * They're generally around 64b.
 307           */
 308          brw->curbe.curbe_bo = drm_intel_bo_alloc(brw->intel.bufmgr, "CURBE",
 309                                                   4096, 1 << 6);
 310          brw->curbe.curbe_next_offset = 0;
 311          drm_intel_gem_bo_map_gtt(brw->curbe.curbe_bo);
 312          assert(bufsz < 4096);
 313       }
 314
 315       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 316       brw->curbe.curbe_next_offset += bufsz;
 317       brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 318
 319       /* Copy data to the buffer:
 320        */
 321       memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset,
 322              buf,
 323              bufsz);
 324    }
 325
 326    /* Because this provokes an action (ie copy the constants into the
 327     * URB), it shouldn't be shortcircuited if identical to the
 328     * previous time - because eg. the urb destination may have
 329     * changed, or the urb contents different to last time.
 330     *
 331     * Note that the data referred to is actually copied internally,
 332     * not just used in place according to passed pointer.
 333     *
 334     * It appears that the CS unit takes care of using each available
 335     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 336     * flushes as necessary when doublebuffering of CURBEs isn't
 337     * possible.
 338     */
 339
 340 emit:
 341    BEGIN_BATCH(2);
 342    if (brw->curbe.total_size == 0) {
 343       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 344       OUT_BATCH(0);
 345    } else {
 346       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 347       OUT_RELOC(brw->curbe.curbe_bo,
 348                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 349                 (brw->curbe.total_size - 1) + brw->curbe.curbe_offset);
 350    }
 351    ADVANCE_BATCH();
 352 }
 353
 354 /* This tracked state is unique in that the state it monitors varies
 355  * dynamically depending on the parameters tracked by the fragment and
 356  * vertex programs.  This is the template used as a starting point,
 357  * each context will maintain a copy of this internally and update as
 358  * required.
 359  */
 360 const struct brw_tracked_state brw_constant_buffer = {
 361    .dirty = {
 362       .mesa = _NEW_PROGRAM_CONSTANTS,
 363       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 364                BRW_NEW_VERTEX_PROGRAM |
 365                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 366                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 367                BRW_NEW_CURBE_OFFSETS |
 368                BRW_NEW_BATCH),
 369       .cache = (CACHE_NEW_WM_PROG)
 370    },
 371    .emit = brw_upload_constant_buffer,
 372 };
 373