src/glsl/link_atomics.cpp

   1 /*
   2  * Copyright © 2013 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
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "glsl_parser_extras.h"
  25 #include "ir.h"
  26 #include "ir_uniform.h"
  27 #include "linker.h"
  28 #include "program/hash_table.h"
  29 #include "main/macros.h"
  30
  31 namespace {
  32    /*
  33     * Atomic counter as seen by the program.
  34     */
  35    struct active_atomic_counter {
  36       unsigned uniform_loc;
  37       ir_variable *var;
  38    };
  39
  40    /*
  41     * Atomic counter buffer referenced by the program.  There is a one
  42     * to one correspondence between these and the objects that can be
  43     * queried using glGetActiveAtomicCounterBufferiv().
  44     */
  45    struct active_atomic_buffer {
  46       active_atomic_buffer()
  47          : counters(0), num_counters(0), stage_references(), size(0)
  48       {}
  49
  50       ~active_atomic_buffer()
  51       {
  52          free(counters);
  53       }
  54
  55       void push_back(unsigned uniform_loc, ir_variable *var)
  56       {
  57          active_atomic_counter *new_counters;
  58
  59          new_counters = (active_atomic_counter *)
  60             realloc(counters, sizeof(active_atomic_counter) *
  61                     (num_counters + 1));
  62
  63          if (new_counters == NULL) {
  64             _mesa_error_no_memory(__func__);
  65             return;
  66          }
  67
  68          counters = new_counters;
  69          counters[num_counters].uniform_loc = uniform_loc;
  70          counters[num_counters].var = var;
  71          num_counters++;
  72       }
  73
  74       active_atomic_counter *counters;
  75       unsigned num_counters;
  76       unsigned stage_references[MESA_SHADER_STAGES];
  77       unsigned size;
  78    };
  79
  80    int
  81    cmp_actives(const void *a, const void *b)
  82    {
  83       const active_atomic_counter *const first = (active_atomic_counter *) a;
  84       const active_atomic_counter *const second = (active_atomic_counter *) b;
  85
  86       return int(first->var->data.atomic.offset) - int(second->var->data.atomic.offset);
  87    }
  88
  89    bool
  90    check_atomic_counters_overlap(const ir_variable *x, const ir_variable *y)
  91    {
  92       return ((x->data.atomic.offset >= y->data.atomic.offset &&
  93                x->data.atomic.offset < y->data.atomic.offset + y->type->atomic_size()) ||
  94               (y->data.atomic.offset >= x->data.atomic.offset &&
  95                y->data.atomic.offset < x->data.atomic.offset + x->type->atomic_size()));
  96    }
  97
  98    void
  99    process_atomic_variable(const glsl_type *t, struct gl_shader_program *prog,
 100                            unsigned *uniform_loc, ir_variable *var,
 101                            active_atomic_buffer *const buffers,
 102                            unsigned *num_buffers, int *offset,
 103                            const unsigned shader_stage)
 104    {
 105       /* FIXME: Arrays of arrays get counted separately. For example:
 106        * x1[3][3][2] = 9 counters
 107        * x2[3][2]    = 3 counters
 108        * x3[2]       = 1 counter
 109        *
 110        * However this code marks all the counters as active even when they
 111        * might not be used.
 112        */
 113       if (t->is_array() && t->fields.array->is_array()) {
 114          for (unsigned i = 0; i < t->length; i++) {
 115             process_atomic_variable(t->fields.array, prog, uniform_loc,
 116                                     var, buffers, num_buffers, offset,
 117                                     shader_stage);
 118          }
 119       } else {
 120          active_atomic_buffer *buf = &buffers[var->data.binding];
 121          gl_uniform_storage *const storage =
 122             &prog->UniformStorage[*uniform_loc];
 123
 124          /* If this is the first time the buffer is used, increment
 125           * the counter of buffers used.
 126           */
 127          if (buf->size == 0)
 128             (*num_buffers)++;
 129
 130          buf->push_back(*uniform_loc, var);
 131
 132          buf->stage_references[shader_stage]++;
 133          buf->size = MAX2(buf->size, *offset + t->atomic_size());
 134
 135          storage->offset = *offset;
 136          *offset += t->atomic_size();
 137
 138          (*uniform_loc)++;
 139       }
 140    }
 141
 142    active_atomic_buffer *
 143    find_active_atomic_counters(struct gl_context *ctx,
 144                                struct gl_shader_program *prog,
 145                                unsigned *num_buffers)
 146    {
 147       active_atomic_buffer *const buffers =
 148          new active_atomic_buffer[ctx->Const.MaxAtomicBufferBindings];
 149
 150       *num_buffers = 0;
 151
 152       for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) {
 153          struct gl_shader *sh = prog->_LinkedShaders[i];
 154          if (sh == NULL)
 155             continue;
 156
 157          foreach_in_list(ir_instruction, node, sh->ir) {
 158             ir_variable *var = node->as_variable();
 159
 160             if (var && var->type->contains_atomic()) {
 161                int offset = var->data.atomic.offset;
 162                unsigned uniform_loc = var->data.location;
 163                process_atomic_variable(var->type, prog, &uniform_loc,
 164                                        var, buffers, num_buffers, &offset, i);
 165             }
 166          }
 167       }
 168
 169       for (unsigned i = 0; i < ctx->Const.MaxAtomicBufferBindings; i++) {
 170          if (buffers[i].size == 0)
 171             continue;
 172
 173          qsort(buffers[i].counters, buffers[i].num_counters,
 174                sizeof(active_atomic_counter),
 175                cmp_actives);
 176
 177          for (unsigned j = 1; j < buffers[i].num_counters; j++) {
 178             /* If an overlapping counter found, it must be a reference to the
 179              * same counter from a different shader stage.
 180              */
 181             if (check_atomic_counters_overlap(buffers[i].counters[j-1].var,
 182                                               buffers[i].counters[j].var)
 183                 && strcmp(buffers[i].counters[j-1].var->name,
 184                           buffers[i].counters[j].var->name) != 0) {
 185                linker_error(prog, "Atomic counter %s declared at offset %d "
 186                             "which is already in use.",
 187                             buffers[i].counters[j].var->name,
 188                             buffers[i].counters[j].var->data.atomic.offset);
 189             }
 190          }
 191       }
 192       return buffers;
 193    }
 194 }
 195
 196 void
 197 link_assign_atomic_counter_resources(struct gl_context *ctx,
 198                                      struct gl_shader_program *prog)
 199 {
 200    unsigned num_buffers;
 201    unsigned num_atomic_buffers[MESA_SHADER_STAGES] = {};
 202    active_atomic_buffer *abs =
 203       find_active_atomic_counters(ctx, prog, &num_buffers);
 204
 205    prog->AtomicBuffers = rzalloc_array(prog, gl_active_atomic_buffer,
 206                                        num_buffers);
 207    prog->NumAtomicBuffers = num_buffers;
 208
 209    unsigned i = 0;
 210    for (unsigned binding = 0;
 211         binding < ctx->Const.MaxAtomicBufferBindings;
 212         binding++) {
 213
 214       /* If the binding was not used, skip.
 215        */
 216       if (abs[binding].size == 0)
 217          continue;
 218
 219       active_atomic_buffer &ab = abs[binding];
 220       gl_active_atomic_buffer &mab = prog->AtomicBuffers[i];
 221
 222       /* Assign buffer-specific fields. */
 223       mab.Binding = binding;
 224       mab.MinimumSize = ab.size;
 225       mab.Uniforms = rzalloc_array(prog->AtomicBuffers, GLuint,
 226                                    ab.num_counters);
 227       mab.NumUniforms = ab.num_counters;
 228
 229       /* Assign counter-specific fields. */
 230       for (unsigned j = 0; j < ab.num_counters; j++) {
 231          ir_variable *const var = ab.counters[j].var;
 232          gl_uniform_storage *const storage =
 233             &prog->UniformStorage[ab.counters[j].uniform_loc];
 234
 235          mab.Uniforms[j] = ab.counters[j].uniform_loc;
 236          if (!var->data.explicit_binding)
 237             var->data.binding = i;
 238
 239          storage->atomic_buffer_index = i;
 240          storage->offset = var->data.atomic.offset;
 241          storage->array_stride = (var->type->is_array() ?
 242                                   var->type->without_array()->atomic_size() : 0);
 243          if (!var->type->is_matrix())
 244             storage->matrix_stride = 0;
 245       }
 246
 247       /* Assign stage-specific fields. */
 248       for (unsigned j = 0; j < MESA_SHADER_STAGES; ++j) {
 249          if (ab.stage_references[j]) {
 250             mab.StageReferences[j] = GL_TRUE;
 251             num_atomic_buffers[j]++;
 252          } else {
 253             mab.StageReferences[j] = GL_FALSE;
 254          }
 255       }
 256
 257       i++;
 258    }
 259
 260    /* Store a list pointers to atomic buffers per stage and store the index
 261     * to the intra-stage buffer list in uniform storage.
 262     */
 263    for (unsigned j = 0; j < MESA_SHADER_STAGES; ++j) {
 264       if (prog->_LinkedShaders[j] && num_atomic_buffers[j] > 0) {
 265          prog->_LinkedShaders[j]->NumAtomicBuffers = num_atomic_buffers[j];
 266          prog->_LinkedShaders[j]->AtomicBuffers =
 267             rzalloc_array(prog, gl_active_atomic_buffer *,
 268                           num_atomic_buffers[j]);
 269
 270          unsigned intra_stage_idx = 0;
 271          for (unsigned i = 0; i < num_buffers; i++) {
 272             struct gl_active_atomic_buffer *atomic_buffer =
 273                &prog->AtomicBuffers[i];
 274             if (atomic_buffer->StageReferences[j]) {
 275                prog->_LinkedShaders[j]->AtomicBuffers[intra_stage_idx] =
 276                   atomic_buffer;
 277
 278                for (unsigned u = 0; u < atomic_buffer->NumUniforms; u++) {
 279                   prog->UniformStorage[atomic_buffer->Uniforms[u]].opaque[j].index =
 280                      intra_stage_idx;
 281                   prog->UniformStorage[atomic_buffer->Uniforms[u]].opaque[j].active =
 282                      true;
 283                }
 284
 285                intra_stage_idx++;
 286             }
 287          }
 288       }
 289    }
 290
 291    delete [] abs;
 292    assert(i == num_buffers);
 293 }
 294
 295 void
 296 link_check_atomic_counter_resources(struct gl_context *ctx,
 297                                     struct gl_shader_program *prog)
 298 {
 299    unsigned num_buffers;
 300    active_atomic_buffer *const abs =
 301       find_active_atomic_counters(ctx, prog, &num_buffers);
 302    unsigned atomic_counters[MESA_SHADER_STAGES] = {};
 303    unsigned atomic_buffers[MESA_SHADER_STAGES] = {};
 304    unsigned total_atomic_counters = 0;
 305    unsigned total_atomic_buffers = 0;
 306
 307    /* Sum the required resources.  Note that this counts buffers and
 308     * counters referenced by several shader stages multiple times
 309     * against the combined limit -- That's the behavior the spec
 310     * requires.
 311     */
 312    for (unsigned i = 0; i < ctx->Const.MaxAtomicBufferBindings; i++) {
 313       if (abs[i].size == 0)
 314          continue;
 315
 316       for (unsigned j = 0; j < MESA_SHADER_STAGES; ++j) {
 317          const unsigned n = abs[i].stage_references[j];
 318
 319          if (n) {
 320             atomic_counters[j] += n;
 321             total_atomic_counters += n;
 322             atomic_buffers[j]++;
 323             total_atomic_buffers++;
 324          }
 325       }
 326    }
 327
 328    /* Check that they are within the supported limits. */
 329    for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) {
 330       if (atomic_counters[i] > ctx->Const.Program[i].MaxAtomicCounters)
 331          linker_error(prog, "Too many %s shader atomic counters",
 332                       _mesa_shader_stage_to_string(i));
 333
 334       if (atomic_buffers[i] > ctx->Const.Program[i].MaxAtomicBuffers)
 335          linker_error(prog, "Too many %s shader atomic counter buffers",
 336                       _mesa_shader_stage_to_string(i));
 337    }
 338
 339    if (total_atomic_counters > ctx->Const.MaxCombinedAtomicCounters)
 340       linker_error(prog, "Too many combined atomic counters");
 341
 342    if (total_atomic_buffers > ctx->Const.MaxCombinedAtomicBuffers)
 343       linker_error(prog, "Too many combined atomic buffers");
 344
 345    delete [] abs;
 346 }