src/gallium/drivers/iris/iris_binder.c

   1 /*
   2  * Copyright © 2018 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 shall be included
  12  * in all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  15  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  20  * DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 /**
  24  * @file iris_binder.c
  25  *
  26  * Shader programs refer to most resources via integer handles.  These are
  27  * indexes (BTIs) into a "Binding Table", which is simply a list of pointers
  28  * to SURFACE_STATE entries.  Each shader stage has its own binding table,
  29  * set by the 3DSTATE_BINDING_TABLE_POINTERS_* commands.  We stream out
  30  * binding tables dynamically, storing them in special BOs we call "binders."
  31  *
  32  * Unfortunately, the hardware designers made 3DSTATE_BINDING_TABLE_POINTERS
  33  * only accept a 16-bit pointer.  This means that all binding tables have to
  34  * live within the 64kB range starting at Surface State Base Address.  (The
  35  * actual SURFACE_STATE entries can live anywhere in the 4GB zone, as the
  36  * binding table entries are full 32-bit pointers.)
  37  *
  38  * To handle this, we split a 4GB region of VMA into two memory zones.
  39  * IRIS_MEMZONE_BINDER is a small region at the bottom able to hold a few
  40  * binder BOs.  IRIS_MEMZONE_SURFACE contains the rest of the 4GB, and is
  41  * always at a higher address than the binders.  This allows us to program
  42  * Surface State Base Address to the binder BO's address, and offset the
  43  * values in the binding table to account for the base not starting at the
  44  * beginning of the 4GB region.
  45  *
  46  * This does mean that we have to emit STATE_BASE_ADDRESS and stall when
  47  * we run out of space in the binder, which hopefully won't happen too often.
  48  */
  49
  50 #include <stdlib.h>
  51 #include "util/u_math.h"
  52 #include "iris_binder.h"
  53 #include "iris_bufmgr.h"
  54 #include "iris_context.h"
  55
  56 #define BTP_ALIGNMENT 32
  57
  58 /* Avoid using offset 0, tools consider it NULL */
  59 #define INIT_INSERT_POINT BTP_ALIGNMENT
  60
  61 static bool
  62 binder_has_space(struct iris_binder *binder, unsigned size)
  63 {
  64    return binder->insert_point + size <= IRIS_BINDER_SIZE;
  65 }
  66
  67 static void
  68 binder_realloc(struct iris_context *ice)
  69 {
  70    struct iris_screen *screen = (void *) ice->ctx.screen;
  71    struct iris_bufmgr *bufmgr = screen->bufmgr;
  72    struct iris_binder *binder = &ice->state.binder;
  73
  74    iris_bo_unreference(binder->bo);
  75
  76    binder->bo =
  77       iris_bo_alloc(bufmgr, "binder", IRIS_BINDER_SIZE, IRIS_MEMZONE_BINDER);
  78    binder->map = iris_bo_map(NULL, binder->bo, MAP_WRITE);
  79    binder->insert_point = INIT_INSERT_POINT;
  80
  81    /* Allocating a new binder requires changing Surface State Base Address,
  82     * which also invalidates all our previous binding tables - each entry
  83     * in those tables is an offset from the old base.
  84     *
  85     * We do this here so that iris_binder_reserve_3d correctly gets a new
  86     * larger total_size when making the updated reservation.
  87     */
  88    ice->state.dirty |= IRIS_ALL_DIRTY_BINDINGS;
  89 }
  90
  91 static uint32_t
  92 binder_insert(struct iris_binder *binder, unsigned size)
  93 {
  94    uint32_t offset = binder->insert_point;
  95
  96    binder->insert_point = align(binder->insert_point + size, BTP_ALIGNMENT);
  97
  98    return offset;
  99 }
 100
 101 /**
 102  * Reserve a block of space in the binder, given the raw size in bytes.
 103  */
 104 uint32_t
 105 iris_binder_reserve(struct iris_context *ice,
 106                     unsigned size)
 107 {
 108    struct iris_binder *binder = &ice->state.binder;
 109
 110    if (!binder_has_space(binder, size))
 111       binder_realloc(ice);
 112
 113    assert(size > 0);
 114    return binder_insert(binder, size);
 115 }
 116
 117 /**
 118  * Reserve and record binder space for 3D pipeline shader stages.
 119  *
 120  * Note that you must actually populate the new binding tables after
 121  * calling this command - the new area is uninitialized.
 122  */
 123 void
 124 iris_binder_reserve_3d(struct iris_context *ice)
 125 {
 126    struct iris_compiled_shader **shaders = ice->shaders.prog;
 127    struct iris_binder *binder = &ice->state.binder;
 128    unsigned sizes[MESA_SHADER_STAGES] = {};
 129    unsigned total_size;
 130
 131    /* If nothing is dirty, skip all this. */
 132    if (!(ice->state.dirty & IRIS_ALL_DIRTY_BINDINGS))
 133       return;
 134
 135    /* Get the binding table sizes for each stage */
 136    for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
 137       if (!shaders[stage])
 138          continue;
 139
 140       const struct brw_stage_prog_data *prog_data =
 141          (const void *) shaders[stage]->prog_data;
 142
 143       /* Round up the size so our next table has an aligned starting offset */
 144       sizes[stage] = align(prog_data->binding_table.size_bytes, BTP_ALIGNMENT);
 145    }
 146
 147    /* Make space for the new binding tables...this may take two tries. */
 148    while (true) {
 149       total_size = 0;
 150       for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
 151          if (ice->state.dirty & (IRIS_DIRTY_BINDINGS_VS << stage))
 152             total_size += sizes[stage];
 153       }
 154
 155       assert(total_size < IRIS_BINDER_SIZE);
 156
 157       if (total_size == 0)
 158          return;
 159
 160       if (binder_has_space(binder, total_size))
 161          break;
 162
 163       /* It didn't fit.  Allocate a new buffer and try again.  Note that
 164        * this will flag all bindings dirty, which may increase total_size
 165        * on the next iteration.
 166        */
 167       binder_realloc(ice);
 168    }
 169
 170    /* Assign space and record the new binding table offsets. */
 171    uint32_t offset = binder_insert(binder, total_size);
 172
 173    for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
 174       if (ice->state.dirty & (IRIS_DIRTY_BINDINGS_VS << stage)) {
 175          binder->bt_offset[stage] = sizes[stage] > 0 ? offset : 0;
 176          offset += sizes[stage];
 177       }
 178    }
 179 }
 180
 181 void
 182 iris_binder_reserve_compute(struct iris_context *ice)
 183 {
 184    if (!(ice->state.dirty & IRIS_DIRTY_BINDINGS_CS))
 185       return;
 186
 187    struct iris_binder *binder = &ice->state.binder;
 188    struct brw_stage_prog_data *prog_data =
 189       ice->shaders.prog[MESA_SHADER_COMPUTE]->prog_data;
 190
 191    unsigned size = prog_data->binding_table.size_bytes;
 192
 193    if (size == 0)
 194       return;
 195
 196    binder->bt_offset[MESA_SHADER_COMPUTE] = iris_binder_reserve(ice, size);
 197 }
 198
 199 void
 200 iris_init_binder(struct iris_context *ice)
 201 {
 202    memset(&ice->state.binder, 0, sizeof(struct iris_binder));
 203    binder_realloc(ice);
 204 }
 205
 206 void
 207 iris_destroy_binder(struct iris_binder *binder)
 208 {
 209    iris_bo_unreference(binder->bo);
 210 }