src/mesa/drivers/dri/i965/gen7_urb.c

   1 /*
   2  * Copyright © 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 "main/macros.h"
  25 #include "intel_batchbuffer.h"
  26 #include "brw_context.h"
  27 #include "brw_state.h"
  28 #include "brw_defines.h"
  29
  30 /**
  31  * The following diagram shows how we partition the URB:
  32  *
  33  *        16kB or 32kB               Rest of the URB space
  34  *   __________-__________   _________________-_________________
  35  *  /                     \ /                                   \
  36  * +-------------------------------------------------------------+
  37  * |     VS/FS/GS Push     |              VS/GS URB              |
  38  * |       Constants       |               Entries               |
  39  * +-------------------------------------------------------------+
  40  *
  41  * Notably, push constants must be stored at the beginning of the URB
  42  * space, while entries can be stored anywhere.  Ivybridge and Haswell
  43  * GT1/GT2 have a maximum constant buffer size of 16kB, while Haswell GT3
  44  * doubles this (32kB).
  45  *
  46  * Ivybridge and Haswell GT1/GT2 allow push constants to be located (and
  47  * sized) in increments of 1kB.  Haswell GT3 requires them to be located and
  48  * sized in increments of 2kB.
  49  *
  50  * Currently we split the constant buffer space evenly among whatever stages
  51  * are active.  This is probably not ideal, but simple.
  52  *
  53  * Ivybridge GT1 and Haswell GT1 have 128kB of URB space.
  54  * Ivybridge GT2 and Haswell GT2 have 256kB of URB space.
  55  * Haswell GT3 has 512kB of URB space.
  56  *
  57  * See "Volume 2a: 3D Pipeline," section 1.8, "Volume 1b: Configurations",
  58  * and the documentation for 3DSTATE_PUSH_CONSTANT_ALLOC_xS.
  59  */
  60 static void
  61 gen7_allocate_push_constants(struct brw_context *brw)
  62 {
  63    unsigned avail_size = 16;
  64    unsigned multiplier =
  65       (brw->gen >= 8 || (brw->is_haswell && brw->gt == 3)) ? 2 : 1;
  66
  67    /* BRW_NEW_GEOMETRY_PROGRAM */
  68    bool gs_present = brw->geometry_program;
  69
  70    unsigned vs_size, gs_size;
  71    if (gs_present) {
  72       vs_size = avail_size / 3;
  73       avail_size -= vs_size;
  74       gs_size = avail_size / 2;
  75       avail_size -= gs_size;
  76    } else {
  77       vs_size = avail_size / 2;
  78       avail_size -= vs_size;
  79       gs_size = 0;
  80    }
  81    unsigned fs_size = avail_size;
  82
  83    gen7_emit_push_constant_state(brw, multiplier * vs_size,
  84                                  multiplier * gs_size, multiplier * fs_size);
  85
  86    /* From p115 of the Ivy Bridge PRM (3.2.1.4 3DSTATE_PUSH_CONSTANT_ALLOC_VS):
  87     *
  88     *     Programming Restriction:
  89     *
  90     *     The 3DSTATE_CONSTANT_VS must be reprogrammed prior to the next
  91     *     3DPRIMITIVE command after programming the
  92     *     3DSTATE_PUSH_CONSTANT_ALLOC_VS.
  93     *
  94     * Similar text exists for the other 3DSTATE_PUSH_CONSTANT_ALLOC_*
  95     * commands.
  96     */
  97    brw->state.dirty.brw |= BRW_NEW_PUSH_CONSTANT_ALLOCATION;
  98 }
  99
 100 void
 101 gen7_emit_push_constant_state(struct brw_context *brw, unsigned vs_size,
 102                               unsigned gs_size, unsigned fs_size)
 103 {
 104    unsigned offset = 0;
 105
 106    BEGIN_BATCH(6);
 107    OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_VS << 16 | (2 - 2));
 108    OUT_BATCH(vs_size | offset << GEN7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
 109    offset += vs_size;
 110
 111    OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_GS << 16 | (2 - 2));
 112    OUT_BATCH(gs_size | offset << GEN7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
 113    offset += gs_size;
 114
 115    OUT_BATCH(_3DSTATE_PUSH_CONSTANT_ALLOC_PS << 16 | (2 - 2));
 116    OUT_BATCH(fs_size | offset << GEN7_PUSH_CONSTANT_BUFFER_OFFSET_SHIFT);
 117    ADVANCE_BATCH();
 118
 119    /* From p292 of the Ivy Bridge PRM (11.2.4 3DSTATE_PUSH_CONSTANT_ALLOC_PS):
 120     *
 121     *     A PIPE_CONTOL command with the CS Stall bit set must be programmed
 122     *     in the ring after this instruction.
 123     *
 124     * No such restriction exists for Haswell or Baytrail.
 125     */
 126    if (brw->gen < 8 && !brw->is_haswell && !brw->is_baytrail)
 127       gen7_emit_cs_stall_flush(brw);
 128 }
 129
 130 const struct brw_tracked_state gen7_push_constant_space = {
 131    .dirty = {
 132       .mesa = 0,
 133       .brw = BRW_NEW_CONTEXT | BRW_NEW_GEOMETRY_PROGRAM,
 134       .cache = 0,
 135    },
 136    .emit = gen7_allocate_push_constants,
 137 };
 138
 139 static void
 140 gen7_upload_urb(struct brw_context *brw)
 141 {
 142    const int push_size_kB =
 143       (brw->gen >= 8 || (brw->is_haswell && brw->gt == 3)) ? 32 : 16;
 144
 145    /* CACHE_NEW_VS_PROG */
 146    unsigned vs_size = MAX2(brw->vs.prog_data->base.urb_entry_size, 1);
 147    unsigned vs_entry_size_bytes = vs_size * 64;
 148    /* BRW_NEW_GEOMETRY_PROGRAM, CACHE_NEW_GS_PROG */
 149    bool gs_present = brw->geometry_program;
 150    unsigned gs_size = gs_present ? brw->gs.prog_data->base.urb_entry_size : 1;
 151    unsigned gs_entry_size_bytes = gs_size * 64;
 152
 153    /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
 154     *
 155     *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
 156     *     Allocation Size is less than 9 512-bit URB entries.
 157     *
 158     * Similar text exists for GS.
 159     */
 160    unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
 161    unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
 162
 163    /* URB allocations must be done in 8k chunks. */
 164    unsigned chunk_size_bytes = 8192;
 165
 166    /* Determine the size of the URB in chunks.
 167     */
 168    unsigned urb_chunks = brw->urb.size * 1024 / chunk_size_bytes;
 169
 170    /* Reserve space for push constants */
 171    unsigned push_constant_bytes = 1024 * push_size_kB;
 172    unsigned push_constant_chunks =
 173       push_constant_bytes / chunk_size_bytes;
 174
 175    /* Initially, assign each stage the minimum amount of URB space it needs,
 176     * and make a note of how much additional space it "wants" (the amount of
 177     * additional space it could actually make use of).
 178     */
 179
 180    /* VS has a lower limit on the number of URB entries */
 181    unsigned vs_chunks =
 182       ALIGN(brw->urb.min_vs_entries * vs_entry_size_bytes, chunk_size_bytes) /
 183       chunk_size_bytes;
 184    unsigned vs_wants =
 185       ALIGN(brw->urb.max_vs_entries * vs_entry_size_bytes,
 186             chunk_size_bytes) / chunk_size_bytes - vs_chunks;
 187
 188    unsigned gs_chunks = 0;
 189    unsigned gs_wants = 0;
 190    if (gs_present) {
 191       /* There are two constraints on the minimum amount of URB space we can
 192        * allocate:
 193        *
 194        * (1) We need room for at least 2 URB entries, since we always operate
 195        * the GS in DUAL_OBJECT mode.
 196        *
 197        * (2) We can't allocate less than nr_gs_entries_granularity.
 198        */
 199       gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
 200                         chunk_size_bytes) / chunk_size_bytes;
 201       gs_wants =
 202          ALIGN(brw->urb.max_gs_entries * gs_entry_size_bytes,
 203                chunk_size_bytes) / chunk_size_bytes - gs_chunks;
 204    }
 205
 206    /* There should always be enough URB space to satisfy the minimum
 207     * requirements of each stage.
 208     */
 209    unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
 210    assert(total_needs <= urb_chunks);
 211
 212    /* Mete out remaining space (if any) in proportion to "wants". */
 213    unsigned total_wants = vs_wants + gs_wants;
 214    unsigned remaining_space = urb_chunks - total_needs;
 215    if (remaining_space > total_wants)
 216       remaining_space = total_wants;
 217    if (remaining_space > 0) {
 218       unsigned vs_additional = (unsigned)
 219          round(vs_wants * (((double) remaining_space) / total_wants));
 220       vs_chunks += vs_additional;
 221       remaining_space -= vs_additional;
 222       gs_chunks += remaining_space;
 223    }
 224
 225    /* Sanity check that we haven't over-allocated. */
 226    assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
 227
 228    /* Finally, compute the number of entries that can fit in the space
 229     * allocated to each stage.
 230     */
 231    unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
 232    unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
 233
 234    /* Since we rounded up when computing *_wants, this may be slightly more
 235     * than the maximum allowed amount, so correct for that.
 236     */
 237    nr_vs_entries = MIN2(nr_vs_entries, brw->urb.max_vs_entries);
 238    nr_gs_entries = MIN2(nr_gs_entries, brw->urb.max_gs_entries);
 239
 240    /* Ensure that we program a multiple of the granularity. */
 241    nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
 242    nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
 243
 244    /* Finally, sanity check to make sure we have at least the minimum number
 245     * of entries needed for each stage.
 246     */
 247    assert(nr_vs_entries >= brw->urb.min_vs_entries);
 248    if (gs_present)
 249       assert(nr_gs_entries >= 2);
 250
 251    /* Gen7 doesn't actually use brw->urb.nr_{vs,gs}_entries, but it seems
 252     * better to put reasonable data in there rather than leave them
 253     * uninitialized.
 254     */
 255    brw->urb.nr_vs_entries = nr_vs_entries;
 256    brw->urb.nr_gs_entries = nr_gs_entries;
 257
 258    /* Lay out the URB in the following order:
 259     * - push constants
 260     * - VS
 261     * - GS
 262     */
 263    brw->urb.vs_start = push_constant_chunks;
 264    brw->urb.gs_start = push_constant_chunks + vs_chunks;
 265
 266    if (brw->gen == 7 && !brw->is_haswell && !brw->is_baytrail)
 267       gen7_emit_vs_workaround_flush(brw);
 268    gen7_emit_urb_state(brw,
 269                        brw->urb.nr_vs_entries, vs_size, brw->urb.vs_start,
 270                        brw->urb.nr_gs_entries, gs_size, brw->urb.gs_start);
 271 }
 272
 273 void
 274 gen7_emit_urb_state(struct brw_context *brw,
 275                     unsigned nr_vs_entries, unsigned vs_size,
 276                     unsigned vs_start, unsigned nr_gs_entries,
 277                     unsigned gs_size, unsigned gs_start)
 278 {
 279    BEGIN_BATCH(8);
 280    OUT_BATCH(_3DSTATE_URB_VS << 16 | (2 - 2));
 281    OUT_BATCH(nr_vs_entries |
 282              ((vs_size - 1) << GEN7_URB_ENTRY_SIZE_SHIFT) |
 283              (vs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 284
 285    OUT_BATCH(_3DSTATE_URB_GS << 16 | (2 - 2));
 286    OUT_BATCH(nr_gs_entries |
 287              ((gs_size - 1) << GEN7_URB_ENTRY_SIZE_SHIFT) |
 288              (gs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 289
 290    /* Allocate the HS and DS zero space - we don't use them. */
 291    OUT_BATCH(_3DSTATE_URB_HS << 16 | (2 - 2));
 292    OUT_BATCH((0 << GEN7_URB_ENTRY_SIZE_SHIFT) |
 293              (vs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 294
 295    OUT_BATCH(_3DSTATE_URB_DS << 16 | (2 - 2));
 296    OUT_BATCH((0 << GEN7_URB_ENTRY_SIZE_SHIFT) |
 297              (vs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 298    ADVANCE_BATCH();
 299 }
 300
 301 const struct brw_tracked_state gen7_urb = {
 302    .dirty = {
 303       .mesa = 0,
 304       .brw = BRW_NEW_CONTEXT | BRW_NEW_GEOMETRY_PROGRAM,
 305       .cache = (CACHE_NEW_VS_PROG | CACHE_NEW_GS_PROG),
 306    },
 307    .emit = gen7_upload_urb,
 308 };