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->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->base.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    BEGIN_BATCH(2);
 150    /* It appears that this is the state packet for the CS unit, ie. the
 151     * urb entries detailed here are housed in the CS range from the
 152     * URB_FENCE command.
 153     */
 154    OUT_BATCH(CMD_CS_URB_STATE << 16 | (2-2));
 155
 156    /* BRW_NEW_URB_FENCE */
 157    if (brw->urb.csize == 0) {
 158       OUT_BATCH(0);
 159    } else {
 160       /* BRW_NEW_URB_FENCE */
 161       assert(brw->urb.nr_cs_entries);
 162       OUT_BATCH((brw->urb.csize - 1) << 4 | brw->urb.nr_cs_entries);
 163    }
 164    CACHED_BATCH();
 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
 181 brw_upload_constant_buffer(struct brw_context *brw)
 182 {
 183    struct gl_context *ctx = &brw->ctx;
 184    const GLuint sz = brw->curbe.total_size;
 185    const GLuint bufsz = sz * 16 * sizeof(GLfloat);
 186    GLfloat *buf;
 187    GLuint i;
 188    gl_clip_plane *clip_planes;
 189
 190    if (sz == 0) {
 191       brw->curbe.last_bufsz  = 0;
 192       goto emit;
 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    /* clipper constants */
 208    if (brw->curbe.clip_size) {
 209       GLuint offset = brw->curbe.clip_start * 16;
 210       GLuint j;
 211
 212       /* If any planes are going this way, send them all this way:
 213        */
 214       for (i = 0; i < 6; i++) {
 215          buf[offset + i * 4 + 0] = fixed_plane[i][0];
 216          buf[offset + i * 4 + 1] = fixed_plane[i][1];
 217          buf[offset + i * 4 + 2] = fixed_plane[i][2];
 218          buf[offset + i * 4 + 3] = fixed_plane[i][3];
 219       }
 220
 221       /* Clip planes: _NEW_TRANSFORM plus _NEW_PROJECTION to get to
 222        * clip-space:
 223        */
 224       clip_planes = brw_select_clip_planes(ctx);
 225       for (j = 0; j < MAX_CLIP_PLANES; j++) {
 226          if (ctx->Transform.ClipPlanesEnabled & (1<<j)) {
 227             buf[offset + i * 4 + 0] = clip_planes[j][0];
 228             buf[offset + i * 4 + 1] = clip_planes[j][1];
 229             buf[offset + i * 4 + 2] = clip_planes[j][2];
 230             buf[offset + i * 4 + 3] = clip_planes[j][3];
 231             i++;
 232          }
 233       }
 234    }
 235
 236    /* vertex shader constants */
 237    if (brw->curbe.vs_size) {
 238       GLuint offset = brw->curbe.vs_start * 16;
 239
 240       for (i = 0; i < brw->vs.prog_data->base.nr_params; i++) {
 241          buf[offset + i] = *brw->vs.prog_data->base.param[i];
 242       }
 243    }
 244
 245    if (0) {
 246       for (i = 0; i < sz*16; i+=4)
 247          printf("curbe %d.%d: %f %f %f %f\n", i/8, i&4,
 248                 buf[i+0], buf[i+1], buf[i+2], buf[i+3]);
 249
 250       printf("last_buf %p buf %p sz %d/%d cmp %d\n",
 251              brw->curbe.last_buf, buf,
 252              bufsz, brw->curbe.last_bufsz,
 253              brw->curbe.last_buf ? memcmp(buf, brw->curbe.last_buf, bufsz) : -1);
 254    }
 255
 256    if (brw->curbe.curbe_bo != NULL &&
 257        bufsz == brw->curbe.last_bufsz &&
 258        memcmp(buf, brw->curbe.last_buf, bufsz) == 0) {
 259       /* constants have not changed */
 260    } else {
 261       /* Update the record of what our last set of constants was.  We
 262        * don't just flip the pointers because we don't fill in the
 263        * data in the padding between the entries.
 264        */
 265       memcpy(brw->curbe.last_buf, buf, bufsz);
 266       brw->curbe.last_bufsz = bufsz;
 267
 268       if (brw->curbe.curbe_bo != NULL &&
 269           brw->curbe.curbe_next_offset + bufsz > brw->curbe.curbe_bo->size)
 270       {
 271          drm_intel_gem_bo_unmap_gtt(brw->curbe.curbe_bo);
 272          drm_intel_bo_unreference(brw->curbe.curbe_bo);
 273          brw->curbe.curbe_bo = NULL;
 274       }
 275
 276       if (brw->curbe.curbe_bo == NULL) {
 277          /* Allocate a single page for CURBE entries for this batchbuffer.
 278           * They're generally around 64b.
 279           */
 280          brw->curbe.curbe_bo = drm_intel_bo_alloc(brw->bufmgr, "CURBE",
 281                                                   4096, 1 << 6);
 282          brw->curbe.curbe_next_offset = 0;
 283          drm_intel_gem_bo_map_gtt(brw->curbe.curbe_bo);
 284          assert(bufsz < 4096);
 285       }
 286
 287       brw->curbe.curbe_offset = brw->curbe.curbe_next_offset;
 288       brw->curbe.curbe_next_offset += bufsz;
 289       brw->curbe.curbe_next_offset = ALIGN(brw->curbe.curbe_next_offset, 64);
 290
 291       /* Copy data to the buffer:
 292        */
 293       memcpy(brw->curbe.curbe_bo->virtual + brw->curbe.curbe_offset,
 294              buf,
 295              bufsz);
 296    }
 297
 298    /* Because this provokes an action (ie copy the constants into the
 299     * URB), it shouldn't be shortcircuited if identical to the
 300     * previous time - because eg. the urb destination may have
 301     * changed, or the urb contents different to last time.
 302     *
 303     * Note that the data referred to is actually copied internally,
 304     * not just used in place according to passed pointer.
 305     *
 306     * It appears that the CS unit takes care of using each available
 307     * URB entry (Const URB Entry == CURBE) in turn, and issuing
 308     * flushes as necessary when doublebuffering of CURBEs isn't
 309     * possible.
 310     */
 311
 312 emit:
 313    BEGIN_BATCH(2);
 314    if (brw->curbe.total_size == 0) {
 315       OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
 316       OUT_BATCH(0);
 317    } else {
 318       OUT_BATCH((CMD_CONST_BUFFER << 16) | (1 << 8) | (2 - 2));
 319       OUT_RELOC(brw->curbe.curbe_bo,
 320                 I915_GEM_DOMAIN_INSTRUCTION, 0,
 321                 (brw->curbe.total_size - 1) + brw->curbe.curbe_offset);
 322    }
 323    ADVANCE_BATCH();
 324 }
 325
 326 const struct brw_tracked_state brw_constant_buffer = {
 327    .dirty = {
 328       .mesa = _NEW_PROGRAM_CONSTANTS,
 329       .brw  = (BRW_NEW_FRAGMENT_PROGRAM |
 330                BRW_NEW_VERTEX_PROGRAM |
 331                BRW_NEW_URB_FENCE | /* Implicit - hardware requires this, not used above */
 332                BRW_NEW_PSP | /* Implicit - hardware requires this, not used above */
 333                BRW_NEW_CURBE_OFFSETS |
 334                BRW_NEW_BATCH),
 335       .cache = (CACHE_NEW_WM_PROG)
 336    },
 337    .emit = brw_upload_constant_buffer,
 338 };
 339