src/compiler/glsl/lower_shared_reference.cpp

   1 /*
   2  * Copyright (c) 2015 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 /**
  25  * \file lower_shared_reference.cpp
  26  *
  27  * IR lower pass to replace dereferences of compute shader shared variables
  28  * with intrinsic function calls.
  29  *
  30  * This relieves drivers of the responsibility of allocating space for the
  31  * shared variables in the shared memory region.
  32  */
  33
  34 #include "lower_buffer_access.h"
  35 #include "ir_builder.h"
  36 #include "main/macros.h"
  37 #include "util/list.h"
  38 #include "glsl_parser_extras.h"
  39
  40 using namespace ir_builder;
  41
  42 namespace {
  43
  44 struct var_offset {
  45    struct list_head node;
  46    const ir_variable *var;
  47    unsigned offset;
  48 };
  49
  50 class lower_shared_reference_visitor :
  51       public lower_buffer_access::lower_buffer_access {
  52 public:
  53
  54    lower_shared_reference_visitor(struct gl_linked_shader *shader)
  55       : list_ctx(ralloc_context(NULL)), shader(shader), shared_size(0u)
  56    {
  57       list_inithead(&var_offsets);
  58    }
  59
  60    ~lower_shared_reference_visitor()
  61    {
  62       ralloc_free(list_ctx);
  63    }
  64
  65    enum {
  66       shared_load_access,
  67       shared_store_access,
  68       shared_atomic_access,
  69    } buffer_access_type;
  70
  71    void insert_buffer_access(void *mem_ctx, ir_dereference *deref,
  72                              const glsl_type *type, ir_rvalue *offset,
  73                              unsigned mask, int channel);
  74
  75    void handle_rvalue(ir_rvalue **rvalue);
  76    ir_visitor_status visit_enter(ir_assignment *ir);
  77    void handle_assignment(ir_assignment *ir);
  78
  79    ir_call *lower_shared_atomic_intrinsic(ir_call *ir);
  80    ir_call *check_for_shared_atomic_intrinsic(ir_call *ir);
  81    ir_visitor_status visit_enter(ir_call *ir);
  82
  83    unsigned get_shared_offset(const ir_variable *);
  84
  85    ir_call *shared_load(void *mem_ctx, const struct glsl_type *type,
  86                         ir_rvalue *offset);
  87    ir_call *shared_store(void *mem_ctx, ir_rvalue *deref, ir_rvalue *offset,
  88                          unsigned write_mask);
  89
  90    void *list_ctx;
  91    struct gl_linked_shader *shader;
  92    struct list_head var_offsets;
  93    unsigned shared_size;
  94    bool progress;
  95 };
  96
  97 unsigned
  98 lower_shared_reference_visitor::get_shared_offset(const ir_variable *var)
  99 {
 100    list_for_each_entry(var_offset, var_entry, &var_offsets, node) {
 101       if (var_entry->var == var)
 102          return var_entry->offset;
 103    }
 104
 105    struct var_offset *new_entry = rzalloc(list_ctx, struct var_offset);
 106    list_add(&new_entry->node, &var_offsets);
 107    new_entry->var = var;
 108
 109    unsigned var_align = var->type->std430_base_alignment(false);
 110    new_entry->offset = glsl_align(shared_size, var_align);
 111
 112    unsigned var_size = var->type->std430_size(false);
 113    shared_size = new_entry->offset + var_size;
 114
 115    return new_entry->offset;
 116 }
 117
 118 void
 119 lower_shared_reference_visitor::handle_rvalue(ir_rvalue **rvalue)
 120 {
 121    if (!*rvalue)
 122       return;
 123
 124    ir_dereference *deref = (*rvalue)->as_dereference();
 125    if (!deref)
 126       return;
 127
 128    ir_variable *var = deref->variable_referenced();
 129    if (!var || var->data.mode != ir_var_shader_shared)
 130       return;
 131
 132    buffer_access_type = shared_load_access;
 133
 134    void *mem_ctx = ralloc_parent(shader->ir);
 135
 136    ir_rvalue *offset = NULL;
 137    unsigned const_offset = get_shared_offset(var);
 138    bool row_major;
 139    int matrix_columns;
 140    assert(var->get_interface_type() == NULL);
 141    const enum glsl_interface_packing packing = GLSL_INTERFACE_PACKING_STD430;
 142
 143    setup_buffer_access(mem_ctx, deref,
 144                        &offset, &const_offset,
 145                        &row_major, &matrix_columns, NULL, packing);
 146
 147    /* Now that we've calculated the offset to the start of the
 148     * dereference, walk over the type and emit loads into a temporary.
 149     */
 150    const glsl_type *type = (*rvalue)->type;
 151    ir_variable *load_var = new(mem_ctx) ir_variable(type,
 152                                                     "shared_load_temp",
 153                                                     ir_var_temporary);
 154    base_ir->insert_before(load_var);
 155
 156    ir_variable *load_offset = new(mem_ctx) ir_variable(glsl_type::uint_type,
 157                                                        "shared_load_temp_offset",
 158                                                        ir_var_temporary);
 159    base_ir->insert_before(load_offset);
 160    base_ir->insert_before(assign(load_offset, offset));
 161
 162    deref = new(mem_ctx) ir_dereference_variable(load_var);
 163
 164    emit_access(mem_ctx, false, deref, load_offset, const_offset, row_major,
 165                matrix_columns, packing, 0);
 166
 167    *rvalue = deref;
 168
 169    progress = true;
 170 }
 171
 172 void
 173 lower_shared_reference_visitor::handle_assignment(ir_assignment *ir)
 174 {
 175    if (!ir || !ir->lhs)
 176       return;
 177
 178    ir_rvalue *rvalue = ir->lhs->as_rvalue();
 179    if (!rvalue)
 180       return;
 181
 182    ir_dereference *deref = ir->lhs->as_dereference();
 183    if (!deref)
 184       return;
 185
 186    ir_variable *var = ir->lhs->variable_referenced();
 187    if (!var || var->data.mode != ir_var_shader_shared)
 188       return;
 189
 190    buffer_access_type = shared_store_access;
 191
 192    /* We have a write to a shared variable, so declare a temporary and rewrite
 193     * the assignment so that the temporary is the LHS.
 194     */
 195    void *mem_ctx = ralloc_parent(shader->ir);
 196
 197    const glsl_type *type = rvalue->type;
 198    ir_variable *store_var = new(mem_ctx) ir_variable(type,
 199                                                      "shared_store_temp",
 200                                                      ir_var_temporary);
 201    base_ir->insert_before(store_var);
 202    ir->lhs = new(mem_ctx) ir_dereference_variable(store_var);
 203
 204    ir_rvalue *offset = NULL;
 205    unsigned const_offset = get_shared_offset(var);
 206    bool row_major;
 207    int matrix_columns;
 208    assert(var->get_interface_type() == NULL);
 209    const enum glsl_interface_packing packing = GLSL_INTERFACE_PACKING_STD430;
 210
 211    setup_buffer_access(mem_ctx, deref,
 212                        &offset, &const_offset,
 213                        &row_major, &matrix_columns, NULL, packing);
 214
 215    deref = new(mem_ctx) ir_dereference_variable(store_var);
 216
 217    ir_variable *store_offset = new(mem_ctx) ir_variable(glsl_type::uint_type,
 218                                                         "shared_store_temp_offset",
 219                                                         ir_var_temporary);
 220    base_ir->insert_before(store_offset);
 221    base_ir->insert_before(assign(store_offset, offset));
 222
 223    /* Now we have to write the value assigned to the temporary back to memory */
 224    emit_access(mem_ctx, true, deref, store_offset, const_offset, row_major,
 225                matrix_columns, packing, ir->write_mask);
 226
 227    progress = true;
 228 }
 229
 230 ir_visitor_status
 231 lower_shared_reference_visitor::visit_enter(ir_assignment *ir)
 232 {
 233    handle_assignment(ir);
 234    return rvalue_visit(ir);
 235 }
 236
 237 void
 238 lower_shared_reference_visitor::insert_buffer_access(void *mem_ctx,
 239                                                      ir_dereference *deref,
 240                                                      const glsl_type *type,
 241                                                      ir_rvalue *offset,
 242                                                      unsigned mask,
 243                                                      int channel)
 244 {
 245    if (buffer_access_type == shared_store_access) {
 246       ir_call *store = shared_store(mem_ctx, deref, offset, mask);
 247       base_ir->insert_after(store);
 248    } else {
 249       ir_call *load = shared_load(mem_ctx, type, offset);
 250       base_ir->insert_before(load);
 251       ir_rvalue *value = load->return_deref->as_rvalue()->clone(mem_ctx, NULL);
 252       base_ir->insert_before(assign(deref->clone(mem_ctx, NULL),
 253                                     value));
 254    }
 255 }
 256
 257 static bool
 258 compute_shader_enabled(const _mesa_glsl_parse_state *state)
 259 {
 260    return state->stage == MESA_SHADER_COMPUTE;
 261 }
 262
 263 ir_call *
 264 lower_shared_reference_visitor::shared_store(void *mem_ctx,
 265                                              ir_rvalue *deref,
 266                                              ir_rvalue *offset,
 267                                              unsigned write_mask)
 268 {
 269    exec_list sig_params;
 270
 271    ir_variable *offset_ref = new(mem_ctx)
 272       ir_variable(glsl_type::uint_type, "offset" , ir_var_function_in);
 273    sig_params.push_tail(offset_ref);
 274
 275    ir_variable *val_ref = new(mem_ctx)
 276       ir_variable(deref->type, "value" , ir_var_function_in);
 277    sig_params.push_tail(val_ref);
 278
 279    ir_variable *writemask_ref = new(mem_ctx)
 280       ir_variable(glsl_type::uint_type, "write_mask" , ir_var_function_in);
 281    sig_params.push_tail(writemask_ref);
 282
 283    ir_function_signature *sig = new(mem_ctx)
 284       ir_function_signature(glsl_type::void_type, compute_shader_enabled);
 285    assert(sig);
 286    sig->replace_parameters(&sig_params);
 287    sig->intrinsic_id = ir_intrinsic_shared_store;
 288
 289    ir_function *f = new(mem_ctx) ir_function("__intrinsic_store_shared");
 290    f->add_signature(sig);
 291
 292    exec_list call_params;
 293    call_params.push_tail(offset->clone(mem_ctx, NULL));
 294    call_params.push_tail(deref->clone(mem_ctx, NULL));
 295    call_params.push_tail(new(mem_ctx) ir_constant(write_mask));
 296    return new(mem_ctx) ir_call(sig, NULL, &call_params);
 297 }
 298
 299 ir_call *
 300 lower_shared_reference_visitor::shared_load(void *mem_ctx,
 301                                             const struct glsl_type *type,
 302                                             ir_rvalue *offset)
 303 {
 304    exec_list sig_params;
 305
 306    ir_variable *offset_ref = new(mem_ctx)
 307       ir_variable(glsl_type::uint_type, "offset_ref" , ir_var_function_in);
 308    sig_params.push_tail(offset_ref);
 309
 310    ir_function_signature *sig =
 311       new(mem_ctx) ir_function_signature(type, compute_shader_enabled);
 312    assert(sig);
 313    sig->replace_parameters(&sig_params);
 314    sig->intrinsic_id = ir_intrinsic_shared_load;
 315
 316    ir_function *f = new(mem_ctx) ir_function("__intrinsic_load_shared");
 317    f->add_signature(sig);
 318
 319    ir_variable *result = new(mem_ctx)
 320       ir_variable(type, "shared_load_result", ir_var_temporary);
 321    base_ir->insert_before(result);
 322    ir_dereference_variable *deref_result = new(mem_ctx)
 323       ir_dereference_variable(result);
 324
 325    exec_list call_params;
 326    call_params.push_tail(offset->clone(mem_ctx, NULL));
 327
 328    return new(mem_ctx) ir_call(sig, deref_result, &call_params);
 329 }
 330
 331 /* Lowers the intrinsic call to a new internal intrinsic that swaps the access
 332  * to the shared variable in the first parameter by an offset. This involves
 333  * creating the new internal intrinsic (i.e. the new function signature).
 334  */
 335 ir_call *
 336 lower_shared_reference_visitor::lower_shared_atomic_intrinsic(ir_call *ir)
 337 {
 338    /* Shared atomics usually have 2 parameters, the shared variable and an
 339     * integer argument. The exception is CompSwap, that has an additional
 340     * integer parameter.
 341     */
 342    int param_count = ir->actual_parameters.length();
 343    assert(param_count == 2 || param_count == 3);
 344
 345    /* First argument must be a scalar integer shared variable */
 346    exec_node *param = ir->actual_parameters.get_head();
 347    ir_instruction *inst = (ir_instruction *) param;
 348    assert(inst->ir_type == ir_type_dereference_variable ||
 349           inst->ir_type == ir_type_dereference_array ||
 350           inst->ir_type == ir_type_dereference_record ||
 351           inst->ir_type == ir_type_swizzle);
 352
 353    ir_rvalue *deref = (ir_rvalue *) inst;
 354    assert(deref->type->is_scalar() && deref->type->is_integer());
 355
 356    ir_variable *var = deref->variable_referenced();
 357    assert(var);
 358
 359    /* Compute the offset to the start if the dereference
 360     */
 361    void *mem_ctx = ralloc_parent(shader->ir);
 362
 363    ir_rvalue *offset = NULL;
 364    unsigned const_offset = get_shared_offset(var);
 365    bool row_major;
 366    int matrix_columns;
 367    assert(var->get_interface_type() == NULL);
 368    const enum glsl_interface_packing packing = GLSL_INTERFACE_PACKING_STD430;
 369    buffer_access_type = shared_atomic_access;
 370
 371    setup_buffer_access(mem_ctx, deref,
 372                        &offset, &const_offset,
 373                        &row_major, &matrix_columns, NULL, packing);
 374
 375    assert(offset);
 376    assert(!row_major);
 377    assert(matrix_columns == 1);
 378
 379    ir_rvalue *deref_offset =
 380       add(offset, new(mem_ctx) ir_constant(const_offset));
 381
 382    /* Create the new internal function signature that will take an offset
 383     * instead of a shared variable
 384     */
 385    exec_list sig_params;
 386    ir_variable *sig_param = new(mem_ctx)
 387       ir_variable(glsl_type::uint_type, "offset" , ir_var_function_in);
 388    sig_params.push_tail(sig_param);
 389
 390    const glsl_type *type = deref->type->base_type == GLSL_TYPE_INT ?
 391       glsl_type::int_type : glsl_type::uint_type;
 392    sig_param = new(mem_ctx)
 393          ir_variable(type, "data1", ir_var_function_in);
 394    sig_params.push_tail(sig_param);
 395
 396    if (param_count == 3) {
 397       sig_param = new(mem_ctx)
 398             ir_variable(type, "data2", ir_var_function_in);
 399       sig_params.push_tail(sig_param);
 400    }
 401
 402    ir_function_signature *sig =
 403       new(mem_ctx) ir_function_signature(deref->type,
 404                                          compute_shader_enabled);
 405    assert(sig);
 406    sig->replace_parameters(&sig_params);
 407
 408    assert(ir->callee->intrinsic_id >= ir_intrinsic_generic_load);
 409    assert(ir->callee->intrinsic_id <= ir_intrinsic_generic_atomic_comp_swap);
 410    sig->intrinsic_id = MAP_INTRINSIC_TO_TYPE(ir->callee->intrinsic_id, shared);
 411
 412    char func_name[64];
 413    sprintf(func_name, "%s_shared", ir->callee_name());
 414    ir_function *f = new(mem_ctx) ir_function(func_name);
 415    f->add_signature(sig);
 416
 417    /* Now, create the call to the internal intrinsic */
 418    exec_list call_params;
 419    call_params.push_tail(deref_offset);
 420    param = ir->actual_parameters.get_head()->get_next();
 421    ir_rvalue *param_as_rvalue = ((ir_instruction *) param)->as_rvalue();
 422    call_params.push_tail(param_as_rvalue->clone(mem_ctx, NULL));
 423    if (param_count == 3) {
 424       param = param->get_next();
 425       param_as_rvalue = ((ir_instruction *) param)->as_rvalue();
 426       call_params.push_tail(param_as_rvalue->clone(mem_ctx, NULL));
 427    }
 428    ir_dereference_variable *return_deref =
 429       ir->return_deref->clone(mem_ctx, NULL);
 430    return new(mem_ctx) ir_call(sig, return_deref, &call_params);
 431 }
 432
 433 ir_call *
 434 lower_shared_reference_visitor::check_for_shared_atomic_intrinsic(ir_call *ir)
 435 {
 436    exec_list& params = ir->actual_parameters;
 437
 438    if (params.length() < 2 || params.length() > 3)
 439       return ir;
 440
 441    ir_rvalue *rvalue =
 442       ((ir_instruction *) params.get_head())->as_rvalue();
 443    if (!rvalue)
 444       return ir;
 445
 446    ir_variable *var = rvalue->variable_referenced();
 447    if (!var || var->data.mode != ir_var_shader_shared)
 448       return ir;
 449
 450    const enum ir_intrinsic_id id = ir->callee->intrinsic_id;
 451    if (id == ir_intrinsic_generic_atomic_add ||
 452        id == ir_intrinsic_generic_atomic_min ||
 453        id == ir_intrinsic_generic_atomic_max ||
 454        id == ir_intrinsic_generic_atomic_and ||
 455        id == ir_intrinsic_generic_atomic_or ||
 456        id == ir_intrinsic_generic_atomic_xor ||
 457        id == ir_intrinsic_generic_atomic_exchange ||
 458        id == ir_intrinsic_generic_atomic_comp_swap) {
 459       return lower_shared_atomic_intrinsic(ir);
 460    }
 461
 462    return ir;
 463 }
 464
 465 ir_visitor_status
 466 lower_shared_reference_visitor::visit_enter(ir_call *ir)
 467 {
 468    ir_call *new_ir = check_for_shared_atomic_intrinsic(ir);
 469    if (new_ir != ir) {
 470       progress = true;
 471       base_ir->replace_with(new_ir);
 472       return visit_continue_with_parent;
 473    }
 474
 475    return rvalue_visit(ir);
 476 }
 477
 478 } /* unnamed namespace */
 479
 480 void
 481 lower_shared_reference(struct gl_linked_shader *shader, unsigned *shared_size)
 482 {
 483    if (shader->Stage != MESA_SHADER_COMPUTE)
 484       return;
 485
 486    lower_shared_reference_visitor v(shader);
 487
 488    /* Loop over the instructions lowering references, because we take a deref
 489     * of an shared variable array using a shared variable dereference as the
 490     * index will produce a collection of instructions all of which have cloned
 491     * shared variable dereferences for that array index.
 492     */
 493    do {
 494       v.progress = false;
 495       visit_list_elements(&v, shader->ir);
 496    } while (v.progress);
 497
 498    *shared_size = v.shared_size;
 499 }