src/gallium/drivers/i965/brw_wm_state.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 #include "util/u_math.h"
  33
  34 #include "brw_context.h"
  35 #include "brw_state.h"
  36 #include "brw_defines.h"
  37 #include "brw_wm.h"
  38 #include "brw_debug.h"
  39 #include "brw_pipe_rast.h"
  40
  41 /***********************************************************************
  42  * WM unit - fragment programs and rasterization
  43  */
  44
  45 struct brw_wm_unit_key {
  46    unsigned int total_grf, total_scratch;
  47    unsigned int urb_entry_read_length;
  48    unsigned int curb_entry_read_length;
  49    unsigned int dispatch_grf_start_reg;
  50
  51    unsigned int curbe_offset;
  52    unsigned int urb_size;
  53
  54    unsigned int max_threads;
  55
  56    unsigned int nr_surfaces, sampler_count;
  57    GLboolean uses_depth, computes_depth, uses_kill, has_flow_control;
  58    GLboolean polygon_stipple, stats_wm, line_stipple, offset_enable;
  59    GLfloat offset_units, offset_factor;
  60 };
  61
  62 static void
  63 wm_unit_populate_key(struct brw_context *brw, struct brw_wm_unit_key *key)
  64 {
  65    const struct brw_fragment_shader *fp = brw->curr.fragment_shader;
  66
  67    memset(key, 0, sizeof(*key));
  68
  69    if (BRW_DEBUG & DEBUG_SINGLE_THREAD)
  70       key->max_threads = 1;
  71    else {
  72       /* WM maximum threads is number of EUs times number of threads per EU. */
  73       if (brw->gen == 5)
  74          key->max_threads = 12 * 6;
  75       else if (brw->is_g4x)
  76          key->max_threads = 10 * 5;
  77       else
  78          key->max_threads = 8 * 4;
  79    }
  80
  81    /* CACHE_NEW_WM_PROG */
  82    key->total_grf = brw->wm.prog_data->total_grf;
  83    key->urb_entry_read_length = brw->wm.prog_data->urb_read_length;
  84    key->curb_entry_read_length = brw->wm.prog_data->curb_read_length;
  85    key->dispatch_grf_start_reg = brw->wm.prog_data->first_curbe_grf;
  86    key->total_scratch = align(brw->wm.prog_data->total_scratch, 1024);
  87
  88    /* BRW_NEW_URB_FENCE */
  89    key->urb_size = brw->urb.vsize;
  90
  91    /* BRW_NEW_CURBE_OFFSETS */
  92    key->curbe_offset = brw->curbe.wm_start;
  93
  94    /* BRW_NEW_NR_SURFACEs */
  95    key->nr_surfaces = brw->wm.nr_surfaces;
  96
  97    /* CACHE_NEW_SAMPLER */
  98    key->sampler_count = brw->wm.sampler_count;
  99
 100    /* PIPE_NEW_RAST */
 101    key->polygon_stipple = brw->curr.rast->templ.poly_stipple_enable;
 102
 103    /* PIPE_NEW_FRAGMENT_PROGRAM */
 104    key->uses_depth = fp->uses_depth;
 105    key->computes_depth = fp->info.writes_z;
 106
 107    /* PIPE_NEW_DEPTH_BUFFER
 108     *
 109     * Override for NULL depthbuffer case, required by the Pixel Shader Computed
 110     * Depth field.
 111     */
 112    if (brw->curr.fb.zsbuf == NULL)
 113       key->computes_depth = 0;
 114
 115    /* PIPE_NEW_DEPTH_STENCIL_ALPHA */
 116    key->uses_kill = (fp->info.uses_kill ||
 117                      brw->curr.zstencil->cc3.alpha_test);
 118
 119    key->has_flow_control = fp->has_flow_control;
 120
 121    /* temporary sanity check assertion */
 122    assert(fp->has_flow_control == 0);
 123
 124    /* PIPE_NEW_QUERY */
 125    key->stats_wm = (brw->query.stats_wm != 0);
 126
 127    /* PIPE_NEW_RAST */
 128    key->line_stipple = brw->curr.rast->templ.line_stipple_enable;
 129
 130
 131    key->offset_enable = (brw->curr.rast->templ.offset_point ||
 132                          brw->curr.rast->templ.offset_line ||
 133                          brw->curr.rast->templ.offset_tri);
 134
 135    key->offset_units = brw->curr.rast->templ.offset_units;
 136    key->offset_factor = brw->curr.rast->templ.offset_scale;
 137 }
 138
 139 /**
 140  * Setup wm hardware state.  See page 225 of Volume 2
 141  */
 142 static enum pipe_error
 143 wm_unit_create_from_key(struct brw_context *brw, struct brw_wm_unit_key *key,
 144                         struct brw_winsys_reloc *reloc,
 145                         unsigned nr_reloc,
 146                         struct brw_winsys_buffer **bo_out)
 147 {
 148    struct brw_wm_unit_state wm;
 149    enum pipe_error ret;
 150
 151    memset(&wm, 0, sizeof(wm));
 152
 153    wm.thread0.grf_reg_count = align(key->total_grf, 16) / 16 - 1;
 154    wm.thread0.kernel_start_pointer = 0; /* reloc */
 155    wm.thread1.depth_coef_urb_read_offset = 1;
 156    wm.thread1.floating_point_mode = BRW_FLOATING_POINT_NON_IEEE_754;
 157
 158    if (brw->gen == 5)
 159       wm.thread1.binding_table_entry_count = 0; /* hardware requirement */
 160    else
 161       wm.thread1.binding_table_entry_count = key->nr_surfaces;
 162
 163    if (key->total_scratch != 0) {
 164       wm.thread2.scratch_space_base_pointer = 0; /* reloc */
 165       wm.thread2.per_thread_scratch_space = key->total_scratch / 1024 - 1;
 166    } else {
 167       wm.thread2.scratch_space_base_pointer = 0;
 168       wm.thread2.per_thread_scratch_space = 0;
 169    }
 170
 171    wm.thread3.dispatch_grf_start_reg = key->dispatch_grf_start_reg;
 172    wm.thread3.urb_entry_read_length = key->urb_entry_read_length;
 173    wm.thread3.urb_entry_read_offset = 0;
 174    wm.thread3.const_urb_entry_read_length = key->curb_entry_read_length;
 175    wm.thread3.const_urb_entry_read_offset = key->curbe_offset * 2;
 176
 177    if (brw->gen == 5)
 178       wm.wm4.sampler_count = 0; /* hardware requirement */
 179    else
 180       wm.wm4.sampler_count = (key->sampler_count + 1) / 4;
 181
 182    /* reloc */
 183    wm.wm4.sampler_state_pointer = 0;
 184
 185    wm.wm5.program_uses_depth = key->uses_depth;
 186    wm.wm5.program_computes_depth = key->computes_depth;
 187    wm.wm5.program_uses_killpixel = key->uses_kill;
 188
 189    if (key->has_flow_control)
 190       wm.wm5.enable_8_pix = 1;
 191    else
 192       wm.wm5.enable_16_pix = 1;
 193
 194    wm.wm5.max_threads = key->max_threads - 1;
 195    wm.wm5.thread_dispatch_enable = 1;   /* AKA: color_write */
 196    wm.wm5.legacy_line_rast = 0;
 197    wm.wm5.legacy_global_depth_bias = 0;
 198    wm.wm5.early_depth_test = 1;         /* never need to disable */
 199    wm.wm5.line_aa_region_width = 0;
 200    wm.wm5.line_endcap_aa_region_width = 1;
 201
 202    wm.wm5.polygon_stipple = key->polygon_stipple;
 203
 204    if (key->offset_enable) {
 205       wm.wm5.depth_offset = 1;
 206       /* Something wierd going on with legacy_global_depth_bias,
 207        * offset_constant, scaling and MRD.  This value passes glean
 208        * but gives some odd results elsewere (eg. the
 209        * quad-offset-units test).
 210        */
 211       wm.global_depth_offset_constant = key->offset_units * 2;
 212
 213       /* This is the only value that passes glean:
 214        */
 215       wm.global_depth_offset_scale = key->offset_factor;
 216    }
 217
 218    wm.wm5.line_stipple = key->line_stipple;
 219
 220    if ((BRW_DEBUG & DEBUG_STATS) || key->stats_wm)
 221       wm.wm4.stats_enable = 1;
 222
 223    ret = brw_upload_cache(&brw->cache, BRW_WM_UNIT,
 224                           key, sizeof(*key),
 225                           reloc, nr_reloc,
 226                           &wm, sizeof(wm),
 227                           NULL, NULL,
 228                           bo_out);
 229    if (ret)
 230       return ret;
 231
 232    return PIPE_OK;
 233 }
 234
 235
 236 static enum pipe_error upload_wm_unit( struct brw_context *brw )
 237 {
 238    struct brw_wm_unit_key key;
 239    struct brw_winsys_reloc reloc[3];
 240    unsigned nr_reloc = 0;
 241    enum pipe_error ret;
 242    unsigned grf_reg_count;
 243    unsigned per_thread_scratch_space;
 244    unsigned stats_enable;
 245    unsigned sampler_count;
 246
 247    wm_unit_populate_key(brw, &key);
 248
 249
 250    /* Allocate the necessary scratch space if we haven't already.  Don't
 251     * bother reducing the allocation later, since we use scratch so
 252     * rarely.
 253     */
 254    assert(key.total_scratch <= 12 * 1024);
 255    if (key.total_scratch) {
 256       GLuint total = key.total_scratch * key.max_threads;
 257
 258       /* Do we need a new buffer:
 259        */
 260       if (brw->wm.scratch_bo && total > brw->wm.scratch_bo->size)
 261          bo_reference(&brw->wm.scratch_bo, NULL);
 262
 263       if (brw->wm.scratch_bo == NULL) {
 264          ret = brw->sws->bo_alloc(brw->sws,
 265                                   BRW_BUFFER_TYPE_SHADER_SCRATCH,
 266                                   total,
 267                                   4096,
 268                                   &brw->wm.scratch_bo);
 269          if (ret)
 270             return ret;
 271       }
 272    }
 273
 274
 275    /* XXX: temporary:
 276     */
 277    grf_reg_count = (align(key.total_grf, 16) / 16 - 1);
 278    per_thread_scratch_space = key.total_scratch / 1024 - 1;
 279    stats_enable = (BRW_DEBUG & DEBUG_STATS) || key.stats_wm;
 280    sampler_count = brw->gen == 5 ? 0 :(key.sampler_count + 1) / 4;
 281
 282    /* Emit WM program relocation */
 283    make_reloc(&reloc[nr_reloc++],
 284               BRW_USAGE_STATE,
 285               grf_reg_count << 1,
 286               offsetof(struct brw_wm_unit_state, thread0),
 287               brw->wm.prog_bo);
 288
 289    /* Emit scratch space relocation */
 290    if (key.total_scratch != 0) {
 291       make_reloc(&reloc[nr_reloc++],
 292                  BRW_USAGE_SCRATCH,
 293                  per_thread_scratch_space,
 294                  offsetof(struct brw_wm_unit_state, thread2),
 295                  brw->wm.scratch_bo);
 296    }
 297
 298    /* Emit sampler state relocation */
 299    if (key.sampler_count != 0) {
 300       make_reloc(&reloc[nr_reloc++],
 301                  BRW_USAGE_STATE,
 302                  stats_enable | (sampler_count << 2),
 303                  offsetof(struct brw_wm_unit_state, wm4),
 304                  brw->wm.sampler_bo);
 305    }
 306
 307
 308    if (brw_search_cache(&brw->cache, BRW_WM_UNIT,
 309                         &key, sizeof(key),
 310                         reloc, nr_reloc,
 311                         NULL,
 312                         &brw->wm.state_bo))
 313       return PIPE_OK;
 314
 315    ret = wm_unit_create_from_key(brw, &key,
 316                                  reloc, nr_reloc,
 317                                  &brw->wm.state_bo);
 318    if (ret)
 319       return ret;
 320
 321    return PIPE_OK;
 322 }
 323
 324 const struct brw_tracked_state brw_wm_unit = {
 325    .dirty = {
 326       .mesa = (PIPE_NEW_FRAGMENT_SHADER |
 327                PIPE_NEW_DEPTH_BUFFER |
 328                PIPE_NEW_RAST |
 329                PIPE_NEW_DEPTH_STENCIL_ALPHA |
 330                PIPE_NEW_QUERY),
 331
 332       .brw = (BRW_NEW_CURBE_OFFSETS |
 333               BRW_NEW_NR_WM_SURFACES),
 334
 335       .cache = (CACHE_NEW_WM_PROG |
 336                 CACHE_NEW_SAMPLER)
 337    },
 338    .prepare = upload_wm_unit,
 339 };
 340