src/intel/common/gen_urb_config.c

   1 /*
   2  * Copyright (c) 2011 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <stdlib.h>
  25 #include <math.h>
  26
  27 #include "util/macros.h"
  28 #include "main/macros.h"
  29 #include "compiler/shader_enums.h"
  30
  31 #include "gen_l3_config.h"
  32
  33 /**
  34  * The following diagram shows how we partition the URB:
  35  *
  36  *        16kb or 32kb               Rest of the URB space
  37  *   __________-__________   _________________-_________________
  38  *  /                     \ /                                   \
  39  * +-------------------------------------------------------------+
  40  * |  VS/HS/DS/GS/FS Push  |           VS/HS/DS/GS URB           |
  41  * |       Constants       |               Entries               |
  42  * +-------------------------------------------------------------+
  43  *
  44  * Push constants must be stored at the beginning of the URB space,
  45  * while URB entries can be stored anywhere.  We choose to lay them
  46  * out in pipeline order (VS -> HS -> DS -> GS).
  47  */
  48
  49 /**
  50  * Decide how to partition the URB among the various stages.
  51  *
  52  * \param[in] push_constant_bytes - space allocate for push constants.
  53  * \param[in] urb_size_bytes - total size of the URB (from L3 config).
  54  * \param[in] tess_present - are tessellation shaders active?
  55  * \param[in] gs_present - are geometry shaders active?
  56  * \param[in] entry_size - the URB entry size (from the shader compiler)
  57  * \param[out] entries - the number of URB entries for each stage
  58  * \param[out] start - the starting offset for each stage
  59  */
  60 void
  61 gen_get_urb_config(const struct gen_device_info *devinfo,
  62                    const struct gen_l3_config *l3_cfg,
  63                    bool tess_present, bool gs_present,
  64                    const unsigned entry_size[4],
  65                    unsigned entries[4], unsigned start[4],
  66                    enum gen_urb_deref_block_size *deref_block_size)
  67 {
  68    unsigned urb_size_kB = gen_get_l3_config_urb_size(devinfo, l3_cfg);
  69
  70    /* RCU_MODE register for Gen12+ in BSpec says:
  71     *
  72     *    "HW reserves 4KB of URB space per bank for Compute Engine out of the
  73     *    total storage available in L3. SW must consider that 4KB of storage
  74     *    per bank will be reduced from what is programmed for the URB space
  75     *    in L3 for Render Engine executed workloads.
  76     *
  77     *    Example: When URB space programmed is 64KB (per bank) for Render
  78     *    Engine, the actual URB space available for operation is only 60KB
  79     *    (per bank). Similarly when URB space programmed is 128KB (per bank)
  80     *    for render engine, the actual URB space available for operation is
  81     *    only 124KB (per bank). More detailed descripton available in "L3
  82     *    Cache" section of the B-Spec."
  83     */
  84    if (devinfo->gen >= 12)
  85       urb_size_kB -= 4 * devinfo->l3_banks;
  86
  87    const unsigned push_constant_kB =
  88       (devinfo->gen >= 8 || (devinfo->is_haswell && devinfo->gt == 3)) ? 32 : 16;
  89
  90    const bool active[4] = { true, tess_present, tess_present, gs_present };
  91
  92    /* URB allocations must be done in 8k chunks. */
  93    const unsigned chunk_size_kB = 8;
  94    const unsigned chunk_size_bytes = chunk_size_kB * 1024;
  95
  96    const unsigned push_constant_chunks = push_constant_kB / chunk_size_kB;
  97    const unsigned urb_chunks = urb_size_kB / chunk_size_kB;
  98
  99    /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
 100     *
 101     *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
 102     *     Allocation Size is less than 9 512-bit URB entries.
 103     *
 104     * Similar text exists for HS, DS and GS.
 105     */
 106    unsigned granularity[4];
 107    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 108       granularity[i] = (entry_size[i] < 9) ? 8 : 1;
 109    }
 110
 111    unsigned min_entries[4] = {
 112       /* VS has a lower limit on the number of URB entries.
 113        *
 114        * From the Broadwell PRM, 3DSTATE_URB_VS instruction:
 115        * "When tessellation is enabled, the VS Number of URB Entries must be
 116        *  greater than or equal to 192."
 117        */
 118       [MESA_SHADER_VERTEX] = tess_present && devinfo->gen == 8 ?
 119          192 : devinfo->urb.min_entries[MESA_SHADER_VERTEX],
 120
 121       /* There are two constraints on the minimum amount of URB space we can
 122        * allocate:
 123        *
 124        * (1) We need room for at least 2 URB entries, since we always operate
 125        * the GS in DUAL_OBJECT mode.
 126        *
 127        * (2) We can't allocate less than nr_gs_entries_granularity.
 128        */
 129       [MESA_SHADER_GEOMETRY] = gs_present ? 2 : 0,
 130
 131       [MESA_SHADER_TESS_CTRL] = tess_present ? 1 : 0,
 132
 133       [MESA_SHADER_TESS_EVAL] = tess_present ?
 134          devinfo->urb.min_entries[MESA_SHADER_TESS_EVAL] : 0,
 135    };
 136
 137    /* Min VS Entries isn't a multiple of 8 on Cherryview/Broxton; round up.
 138     * Round them all up.
 139     */
 140    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 141       min_entries[i] = ALIGN(min_entries[i], granularity[i]);
 142    }
 143
 144    unsigned entry_size_bytes[4];
 145    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 146       entry_size_bytes[i] = 64 * entry_size[i];
 147    }
 148
 149    /* Initially, assign each stage the minimum amount of URB space it needs,
 150     * and make a note of how much additional space it "wants" (the amount of
 151     * additional space it could actually make use of).
 152     */
 153    unsigned chunks[4];
 154    unsigned wants[4];
 155    unsigned total_needs = push_constant_chunks;
 156    unsigned total_wants = 0;
 157
 158    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 159       if (active[i]) {
 160          chunks[i] = DIV_ROUND_UP(min_entries[i] * entry_size_bytes[i],
 161                                   chunk_size_bytes);
 162
 163          wants[i] =
 164             DIV_ROUND_UP(devinfo->urb.max_entries[i] * entry_size_bytes[i],
 165                          chunk_size_bytes) - chunks[i];
 166       } else {
 167          chunks[i] = 0;
 168          wants[i] = 0;
 169       }
 170
 171       total_needs += chunks[i];
 172       total_wants += wants[i];
 173    }
 174
 175    assert(total_needs <= urb_chunks);
 176
 177    /* Mete out remaining space (if any) in proportion to "wants". */
 178    unsigned remaining_space = MIN2(urb_chunks - total_needs, total_wants);
 179
 180    if (remaining_space > 0) {
 181       for (int i = MESA_SHADER_VERTEX;
 182            total_wants > 0 && i <= MESA_SHADER_TESS_EVAL; i++) {
 183          unsigned additional = (unsigned)
 184             roundf(wants[i] * (((float) remaining_space) / total_wants));
 185          chunks[i] += additional;
 186          remaining_space -= additional;
 187          total_wants -= wants[i];
 188       }
 189
 190       chunks[MESA_SHADER_GEOMETRY] += remaining_space;
 191    }
 192
 193    /* Sanity check that we haven't over-allocated. */
 194    unsigned total_chunks = push_constant_chunks;
 195    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 196       total_chunks += chunks[i];
 197    }
 198    assert(total_chunks <= urb_chunks);
 199
 200    /* Finally, compute the number of entries that can fit in the space
 201     * allocated to each stage.
 202     */
 203    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 204       entries[i] = chunks[i] * chunk_size_bytes / entry_size_bytes[i];
 205
 206       /* Since we rounded up when computing wants[], this may be slightly
 207        * more than the maximum allowed amount, so correct for that.
 208        */
 209       entries[i] = MIN2(entries[i], devinfo->urb.max_entries[i]);
 210
 211       /* Ensure that we program a multiple of the granularity. */
 212       entries[i] = ROUND_DOWN_TO(entries[i], granularity[i]);
 213
 214       /* Finally, sanity check to make sure we have at least the minimum
 215        * number of entries needed for each stage.
 216        */
 217       assert(entries[i] >= min_entries[i]);
 218    }
 219
 220    /* Lay out the URB in pipeline order: push constants, VS, HS, DS, GS. */
 221    int next = push_constant_chunks;
 222    for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
 223       if (entries[i]) {
 224          start[i] = next;
 225          next += chunks[i];
 226       } else {
 227          /* Just put disabled stages at the beginning. */
 228          start[i] = 0;
 229       }
 230    }
 231
 232    if (deref_block_size) {
 233       if (devinfo->gen >= 12) {
 234          /* From the Gen12 BSpec:
 235           *
 236           *    "Deref Block size depends on the last enabled shader and number
 237           *    of handles programmed for that shader
 238           *
 239           *       1) For GS last shader enabled cases, the deref block is
 240           *          always set to a per poly(within hardware)
 241           *
 242           *    If the last enabled shader is VS or DS.
 243           *
 244           *       1) If DS is last enabled shader then if the number of DS
 245           *          handles is less than 324, need to set per poly deref.
 246           *
 247           *       2) If VS is last enabled shader then if the number of VS
 248           *          handles is less than 192, need to set per poly deref"
 249           *
 250           * The default is 32 so we assume that's the right choice if we're
 251           * not in one of the explicit cases listed above.
 252           */
 253          if (gs_present) {
 254             *deref_block_size = GEN_URB_DEREF_BLOCK_SIZE_PER_POLY;
 255          } else if (tess_present) {
 256             if (entries[MESA_SHADER_TESS_EVAL] < 324)
 257                *deref_block_size = GEN_URB_DEREF_BLOCK_SIZE_PER_POLY;
 258             else
 259                *deref_block_size = GEN_URB_DEREF_BLOCK_SIZE_32;
 260          } else {
 261             if (entries[MESA_SHADER_VERTEX] < 192)
 262                *deref_block_size = GEN_URB_DEREF_BLOCK_SIZE_PER_POLY;
 263             else
 264                *deref_block_size = GEN_URB_DEREF_BLOCK_SIZE_32;
 265          }
 266       } else {
 267          *deref_block_size = 0;
 268       }
 269    }
 270 }