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->ctx.NewDriverState |= 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    },
 135    .emit = gen7_allocate_push_constants,
 136 };
 137
 138 static void
 139 gen7_upload_urb(struct brw_context *brw)
 140 {
 141    const int push_size_kB =
 142       (brw->gen >= 8 || (brw->is_haswell && brw->gt == 3)) ? 32 : 16;
 143
 144    /* BRW_NEW_VS_PROG_DATA */
 145    unsigned vs_size = MAX2(brw->vs.prog_data->base.urb_entry_size, 1);
 146    unsigned vs_entry_size_bytes = vs_size * 64;
 147    /* BRW_NEW_GEOMETRY_PROGRAM, BRW_NEW_GS_PROG_DATA */
 148    bool gs_present = brw->geometry_program;
 149    unsigned gs_size = gs_present ? brw->gs.prog_data->base.urb_entry_size : 1;
 150    unsigned gs_entry_size_bytes = gs_size * 64;
 151
 152    /* If we're just switching between programs with the same URB requirements,
 153     * skip the rest of the logic.
 154     */
 155    if (!(brw->ctx.NewDriverState & BRW_NEW_CONTEXT) &&
 156        brw->urb.vsize == vs_size &&
 157        brw->urb.gs_present == gs_present &&
 158        brw->urb.gsize == gs_size) {
 159       return;
 160    }
 161    brw->urb.vsize = vs_size;
 162    brw->urb.gs_present = gs_present;
 163    brw->urb.gsize = gs_size;
 164
 165    /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
 166     *
 167     *     VS Number of URB Entries must be divisible by 8 if the VS URB Entry
 168     *     Allocation Size is less than 9 512-bit URB entries.
 169     *
 170     * Similar text exists for GS.
 171     */
 172    unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
 173    unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
 174
 175    /* URB allocations must be done in 8k chunks. */
 176    unsigned chunk_size_bytes = 8192;
 177
 178    /* Determine the size of the URB in chunks.
 179     */
 180    unsigned urb_chunks = brw->urb.size * 1024 / chunk_size_bytes;
 181
 182    /* Reserve space for push constants */
 183    unsigned push_constant_bytes = 1024 * push_size_kB;
 184    unsigned push_constant_chunks =
 185       push_constant_bytes / chunk_size_bytes;
 186
 187    /* Initially, assign each stage the minimum amount of URB space it needs,
 188     * and make a note of how much additional space it "wants" (the amount of
 189     * additional space it could actually make use of).
 190     */
 191
 192    /* VS has a lower limit on the number of URB entries */
 193    unsigned vs_chunks =
 194       ALIGN(brw->urb.min_vs_entries * vs_entry_size_bytes, chunk_size_bytes) /
 195       chunk_size_bytes;
 196    unsigned vs_wants =
 197       ALIGN(brw->urb.max_vs_entries * vs_entry_size_bytes,
 198             chunk_size_bytes) / chunk_size_bytes - vs_chunks;
 199
 200    unsigned gs_chunks = 0;
 201    unsigned gs_wants = 0;
 202    if (gs_present) {
 203       /* There are two constraints on the minimum amount of URB space we can
 204        * allocate:
 205        *
 206        * (1) We need room for at least 2 URB entries, since we always operate
 207        * the GS in DUAL_OBJECT mode.
 208        *
 209        * (2) We can't allocate less than nr_gs_entries_granularity.
 210        */
 211       gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
 212                         chunk_size_bytes) / chunk_size_bytes;
 213       gs_wants =
 214          ALIGN(brw->urb.max_gs_entries * gs_entry_size_bytes,
 215                chunk_size_bytes) / chunk_size_bytes - gs_chunks;
 216    }
 217
 218    /* There should always be enough URB space to satisfy the minimum
 219     * requirements of each stage.
 220     */
 221    unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
 222    assert(total_needs <= urb_chunks);
 223
 224    /* Mete out remaining space (if any) in proportion to "wants". */
 225    unsigned total_wants = vs_wants + gs_wants;
 226    unsigned remaining_space = urb_chunks - total_needs;
 227    if (remaining_space > total_wants)
 228       remaining_space = total_wants;
 229    if (remaining_space > 0) {
 230       unsigned vs_additional = (unsigned)
 231          roundf(vs_wants * (((float) remaining_space) / total_wants));
 232       vs_chunks += vs_additional;
 233       remaining_space -= vs_additional;
 234       gs_chunks += remaining_space;
 235    }
 236
 237    /* Sanity check that we haven't over-allocated. */
 238    assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
 239
 240    /* Finally, compute the number of entries that can fit in the space
 241     * allocated to each stage.
 242     */
 243    unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
 244    unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
 245
 246    /* Since we rounded up when computing *_wants, this may be slightly more
 247     * than the maximum allowed amount, so correct for that.
 248     */
 249    nr_vs_entries = MIN2(nr_vs_entries, brw->urb.max_vs_entries);
 250    nr_gs_entries = MIN2(nr_gs_entries, brw->urb.max_gs_entries);
 251
 252    /* Ensure that we program a multiple of the granularity. */
 253    nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
 254    nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
 255
 256    /* Finally, sanity check to make sure we have at least the minimum number
 257     * of entries needed for each stage.
 258     */
 259    assert(nr_vs_entries >= brw->urb.min_vs_entries);
 260    if (gs_present)
 261       assert(nr_gs_entries >= 2);
 262
 263    /* Gen7 doesn't actually use brw->urb.nr_{vs,gs}_entries, but it seems
 264     * better to put reasonable data in there rather than leave them
 265     * uninitialized.
 266     */
 267    brw->urb.nr_vs_entries = nr_vs_entries;
 268    brw->urb.nr_gs_entries = nr_gs_entries;
 269
 270    /* Lay out the URB in the following order:
 271     * - push constants
 272     * - VS
 273     * - GS
 274     */
 275    brw->urb.vs_start = push_constant_chunks;
 276    brw->urb.gs_start = push_constant_chunks + vs_chunks;
 277
 278    if (brw->gen == 7 && !brw->is_haswell && !brw->is_baytrail)
 279       gen7_emit_vs_workaround_flush(brw);
 280    gen7_emit_urb_state(brw,
 281                        brw->urb.nr_vs_entries, vs_size, brw->urb.vs_start,
 282                        brw->urb.nr_gs_entries, gs_size, brw->urb.gs_start);
 283 }
 284
 285 void
 286 gen7_emit_urb_state(struct brw_context *brw,
 287                     unsigned nr_vs_entries, unsigned vs_size,
 288                     unsigned vs_start, unsigned nr_gs_entries,
 289                     unsigned gs_size, unsigned gs_start)
 290 {
 291    BEGIN_BATCH(8);
 292    OUT_BATCH(_3DSTATE_URB_VS << 16 | (2 - 2));
 293    OUT_BATCH(nr_vs_entries |
 294              ((vs_size - 1) << GEN7_URB_ENTRY_SIZE_SHIFT) |
 295              (vs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 296
 297    OUT_BATCH(_3DSTATE_URB_GS << 16 | (2 - 2));
 298    OUT_BATCH(nr_gs_entries |
 299              ((gs_size - 1) << GEN7_URB_ENTRY_SIZE_SHIFT) |
 300              (gs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 301
 302    /* Allocate the HS and DS zero space - we don't use them. */
 303    OUT_BATCH(_3DSTATE_URB_HS << 16 | (2 - 2));
 304    OUT_BATCH((0 << GEN7_URB_ENTRY_SIZE_SHIFT) |
 305              (vs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 306
 307    OUT_BATCH(_3DSTATE_URB_DS << 16 | (2 - 2));
 308    OUT_BATCH((0 << GEN7_URB_ENTRY_SIZE_SHIFT) |
 309              (vs_start << GEN7_URB_STARTING_ADDRESS_SHIFT));
 310    ADVANCE_BATCH();
 311 }
 312
 313 const struct brw_tracked_state gen7_urb = {
 314    .dirty = {
 315       .mesa = 0,
 316       .brw = BRW_NEW_CONTEXT |
 317              BRW_NEW_GEOMETRY_PROGRAM |
 318              BRW_NEW_GS_PROG_DATA |
 319              BRW_NEW_VS_PROG_DATA,
 320    },
 321    .emit = gen7_upload_urb,
 322 };