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