src/gallium/drivers/radeonsi/si_shader_nir.c

   1 /*
   2  * Copyright 2017 Advanced Micro Devices, Inc.
   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  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include "si_shader.h"
  25 #include "si_shader_internal.h"
  26
  27 #include "ac_nir_to_llvm.h"
  28
  29 #include "tgsi/tgsi_from_mesa.h"
  30
  31 #include "compiler/nir/nir.h"
  32 #include "compiler/nir_types.h"
  33
  34
  35 static int
  36 type_size(const struct glsl_type *type)
  37 {
  38    return glsl_count_attribute_slots(type, false);
  39 }
  40
  41 static void scan_instruction(struct tgsi_shader_info *info,
  42                              nir_instr *instr)
  43 {
  44         if (instr->type == nir_instr_type_alu) {
  45                 nir_alu_instr *alu = nir_instr_as_alu(instr);
  46
  47                 switch (alu->op) {
  48                 case nir_op_fddx:
  49                 case nir_op_fddy:
  50                 case nir_op_fddx_fine:
  51                 case nir_op_fddy_fine:
  52                 case nir_op_fddx_coarse:
  53                 case nir_op_fddy_coarse:
  54                         info->uses_derivatives = true;
  55                         break;
  56                 default:
  57                         break;
  58                 }
  59         } else if (instr->type == nir_instr_type_tex) {
  60                 nir_tex_instr *tex = nir_instr_as_tex(instr);
  61
  62                 if (!tex->texture) {
  63                         info->samplers_declared |=
  64                                 u_bit_consecutive(tex->sampler_index, 1);
  65                 }
  66
  67                 switch (tex->op) {
  68                 case nir_texop_tex:
  69                 case nir_texop_txb:
  70                 case nir_texop_lod:
  71                         info->uses_derivatives = true;
  72                         break;
  73                 default:
  74                         break;
  75                 }
  76         } else if (instr->type == nir_instr_type_intrinsic) {
  77                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
  78
  79                 switch (intr->intrinsic) {
  80                 case nir_intrinsic_load_front_face:
  81                         info->uses_frontface = 1;
  82                         break;
  83                 case nir_intrinsic_load_instance_id:
  84                         info->uses_instanceid = 1;
  85                         break;
  86                 case nir_intrinsic_load_vertex_id:
  87                         info->uses_vertexid = 1;
  88                         break;
  89                 case nir_intrinsic_load_vertex_id_zero_base:
  90                         info->uses_vertexid_nobase = 1;
  91                         break;
  92                 case nir_intrinsic_load_base_vertex:
  93                         info->uses_basevertex = 1;
  94                         break;
  95                 case nir_intrinsic_load_primitive_id:
  96                         info->uses_primid = 1;
  97                         break;
  98                 case nir_intrinsic_image_store:
  99                 case nir_intrinsic_image_atomic_add:
 100                 case nir_intrinsic_image_atomic_min:
 101                 case nir_intrinsic_image_atomic_max:
 102                 case nir_intrinsic_image_atomic_and:
 103                 case nir_intrinsic_image_atomic_or:
 104                 case nir_intrinsic_image_atomic_xor:
 105                 case nir_intrinsic_image_atomic_exchange:
 106                 case nir_intrinsic_image_atomic_comp_swap:
 107                 case nir_intrinsic_store_ssbo:
 108                 case nir_intrinsic_ssbo_atomic_add:
 109                 case nir_intrinsic_ssbo_atomic_imin:
 110                 case nir_intrinsic_ssbo_atomic_umin:
 111                 case nir_intrinsic_ssbo_atomic_imax:
 112                 case nir_intrinsic_ssbo_atomic_umax:
 113                 case nir_intrinsic_ssbo_atomic_and:
 114                 case nir_intrinsic_ssbo_atomic_or:
 115                 case nir_intrinsic_ssbo_atomic_xor:
 116                 case nir_intrinsic_ssbo_atomic_exchange:
 117                 case nir_intrinsic_ssbo_atomic_comp_swap:
 118                         info->writes_memory = true;
 119                         break;
 120                 default:
 121                         break;
 122                 }
 123         }
 124 }
 125
 126 void si_nir_scan_shader(const struct nir_shader *nir,
 127                         struct tgsi_shader_info *info)
 128 {
 129         nir_function *func;
 130         unsigned i;
 131
 132         assert(nir->info.stage == MESA_SHADER_VERTEX ||
 133                nir->info.stage == MESA_SHADER_FRAGMENT);
 134
 135         info->processor = pipe_shader_type_from_mesa(nir->info.stage);
 136         info->num_tokens = 2; /* indicate that the shader is non-empty */
 137         info->num_instructions = 2;
 138
 139         info->num_inputs = nir->num_inputs;
 140         info->num_outputs = nir->num_outputs;
 141
 142         i = 0;
 143         nir_foreach_variable(variable, &nir->inputs) {
 144                 unsigned semantic_name, semantic_index;
 145                 unsigned attrib_count = glsl_count_attribute_slots(variable->type,
 146                                                                    nir->info.stage == MESA_SHADER_VERTEX);
 147
 148                 assert(attrib_count == 1 && "not implemented");
 149
 150                 /* Vertex shader inputs don't have semantics. The state
 151                  * tracker has already mapped them to attributes via
 152                  * variable->data.driver_location.
 153                  */
 154                 if (nir->info.stage == MESA_SHADER_VERTEX)
 155                         continue;
 156
 157                 /* Fragment shader position is a system value. */
 158                 if (nir->info.stage == MESA_SHADER_FRAGMENT &&
 159                     variable->data.location == VARYING_SLOT_POS) {
 160                         if (variable->data.pixel_center_integer)
 161                                 info->properties[TGSI_PROPERTY_FS_COORD_PIXEL_CENTER] =
 162                                         TGSI_FS_COORD_PIXEL_CENTER_INTEGER;
 163                         continue;
 164                 }
 165
 166                 tgsi_get_gl_varying_semantic(variable->data.location, true,
 167                                              &semantic_name, &semantic_index);
 168
 169                 info->input_semantic_name[i] = semantic_name;
 170                 info->input_semantic_index[i] = semantic_index;
 171
 172                 if (variable->data.sample)
 173                         info->input_interpolate_loc[i] = TGSI_INTERPOLATE_LOC_SAMPLE;
 174                 else if (variable->data.centroid)
 175                         info->input_interpolate_loc[i] = TGSI_INTERPOLATE_LOC_CENTROID;
 176                 else
 177                         info->input_interpolate_loc[i] = TGSI_INTERPOLATE_LOC_CENTER;
 178
 179                 enum glsl_base_type base_type =
 180                         glsl_get_base_type(glsl_without_array(variable->type));
 181
 182                 switch (variable->data.interpolation) {
 183                 case INTERP_MODE_NONE:
 184                         if (glsl_base_type_is_integer(base_type)) {
 185                                 info->input_interpolate[i] = TGSI_INTERPOLATE_CONSTANT;
 186                                 break;
 187                         }
 188
 189                         if (semantic_name == TGSI_SEMANTIC_COLOR) {
 190                                 info->input_interpolate[i] = TGSI_INTERPOLATE_COLOR;
 191                                 goto persp_locations;
 192                         }
 193                         /* fall-through */
 194                 case INTERP_MODE_SMOOTH:
 195                         assert(!glsl_base_type_is_integer(base_type));
 196
 197                         info->input_interpolate[i] = TGSI_INTERPOLATE_PERSPECTIVE;
 198
 199                 persp_locations:
 200                         if (variable->data.sample)
 201                                 info->uses_persp_sample = true;
 202                         else if (variable->data.centroid)
 203                                 info->uses_persp_centroid = true;
 204                         else
 205                                 info->uses_persp_center = true;
 206                         break;
 207
 208                 case INTERP_MODE_NOPERSPECTIVE:
 209                         assert(!glsl_base_type_is_integer(base_type));
 210
 211                         info->input_interpolate[i] = TGSI_INTERPOLATE_LINEAR;
 212
 213                         if (variable->data.sample)
 214                                 info->uses_linear_sample = true;
 215                         else if (variable->data.centroid)
 216                                 info->uses_linear_centroid = true;
 217                         else
 218                                 info->uses_linear_center = true;
 219                         break;
 220
 221                 case INTERP_MODE_FLAT:
 222                         info->input_interpolate[i] = TGSI_INTERPOLATE_CONSTANT;
 223                         break;
 224                 }
 225
 226                 /* TODO make this more precise */
 227                 if (variable->data.location == VARYING_SLOT_COL0)
 228                         info->colors_read |= 0x0f;
 229                 else if (variable->data.location == VARYING_SLOT_COL1)
 230                         info->colors_read |= 0xf0;
 231
 232                 i++;
 233         }
 234
 235         i = 0;
 236         nir_foreach_variable(variable, &nir->outputs) {
 237                 unsigned semantic_name, semantic_index;
 238
 239                 if (nir->info.stage == MESA_SHADER_FRAGMENT) {
 240                         tgsi_get_gl_frag_result_semantic(variable->data.location,
 241                                 &semantic_name, &semantic_index);
 242                 } else {
 243                         tgsi_get_gl_varying_semantic(variable->data.location, true,
 244                                                      &semantic_name, &semantic_index);
 245                 }
 246
 247                 info->output_semantic_name[i] = semantic_name;
 248                 info->output_semantic_index[i] = semantic_index;
 249                 info->output_usagemask[i] = TGSI_WRITEMASK_XYZW;
 250
 251                 unsigned num_components = 4;
 252                 unsigned vector_elements = glsl_get_vector_elements(glsl_without_array(variable->type));
 253                 if (vector_elements)
 254                         num_components = vector_elements;
 255
 256                 unsigned gs_out_streams;
 257                 if (variable->data.stream & (1u << 31)) {
 258                         gs_out_streams = variable->data.stream & ~(1u << 31);
 259                 } else {
 260                         assert(variable->data.stream < 4);
 261                         gs_out_streams = 0;
 262                         for (unsigned j = 0; j < num_components; ++j)
 263                                 gs_out_streams |= variable->data.stream << (2 * (variable->data.location_frac + j));
 264                 }
 265
 266                 unsigned streamx = gs_out_streams & 3;
 267                 unsigned streamy = (gs_out_streams >> 2) & 3;
 268                 unsigned streamz = (gs_out_streams >> 4) & 3;
 269                 unsigned streamw = (gs_out_streams >> 6) & 3;
 270
 271                 if (info->output_usagemask[i] & TGSI_WRITEMASK_X) {
 272                         info->output_streams[i] |= streamx;
 273                         info->num_stream_output_components[streamx]++;
 274                 }
 275                 if (info->output_usagemask[i] & TGSI_WRITEMASK_Y) {
 276                         info->output_streams[i] |= streamy << 2;
 277                         info->num_stream_output_components[streamy]++;
 278                 }
 279                 if (info->output_usagemask[i] & TGSI_WRITEMASK_Z) {
 280                         info->output_streams[i] |= streamz << 4;
 281                         info->num_stream_output_components[streamz]++;
 282                 }
 283                 if (info->output_usagemask[i] & TGSI_WRITEMASK_W) {
 284                         info->output_streams[i] |= streamw << 6;
 285                         info->num_stream_output_components[streamw]++;
 286                 }
 287
 288                 switch (semantic_name) {
 289                 case TGSI_SEMANTIC_PRIMID:
 290                         info->writes_primid = true;
 291                         break;
 292                 case TGSI_SEMANTIC_VIEWPORT_INDEX:
 293                         info->writes_viewport_index = true;
 294                         break;
 295                 case TGSI_SEMANTIC_LAYER:
 296                         info->writes_layer = true;
 297                         break;
 298                 case TGSI_SEMANTIC_PSIZE:
 299                         info->writes_psize = true;
 300                         break;
 301                 case TGSI_SEMANTIC_CLIPVERTEX:
 302                         info->writes_clipvertex = true;
 303                         break;
 304                 case TGSI_SEMANTIC_COLOR:
 305                         info->colors_written |= 1 << semantic_index;
 306                         break;
 307                 case TGSI_SEMANTIC_STENCIL:
 308                         info->writes_stencil = true;
 309                         break;
 310                 case TGSI_SEMANTIC_SAMPLEMASK:
 311                         info->writes_samplemask = true;
 312                         break;
 313                 case TGSI_SEMANTIC_EDGEFLAG:
 314                         info->writes_edgeflag = true;
 315                         break;
 316                 case TGSI_SEMANTIC_POSITION:
 317                         if (info->processor == PIPE_SHADER_FRAGMENT)
 318                                 info->writes_z = true;
 319                         else
 320                                 info->writes_position = true;
 321                         break;
 322                 }
 323
 324                 i++;
 325         }
 326
 327         nir_foreach_variable(variable, &nir->uniforms) {
 328                 const struct glsl_type *type = variable->type;
 329                 enum glsl_base_type base_type =
 330                         glsl_get_base_type(glsl_without_array(type));
 331                 unsigned aoa_size = MAX2(1, glsl_get_aoa_size(type));
 332
 333                 /* We rely on the fact that nir_lower_samplers_as_deref has
 334                  * eliminated struct dereferences.
 335                  */
 336                 if (base_type == GLSL_TYPE_SAMPLER)
 337                         info->samplers_declared |=
 338                                 u_bit_consecutive(variable->data.binding, aoa_size);
 339                 else if (base_type == GLSL_TYPE_IMAGE)
 340                         info->images_declared |=
 341                                 u_bit_consecutive(variable->data.binding, aoa_size);
 342         }
 343
 344         info->num_written_clipdistance = nir->info.clip_distance_array_size;
 345         info->num_written_culldistance = nir->info.cull_distance_array_size;
 346         info->clipdist_writemask = u_bit_consecutive(0, info->num_written_clipdistance);
 347         info->culldist_writemask = u_bit_consecutive(0, info->num_written_culldistance);
 348
 349         if (info->processor == PIPE_SHADER_FRAGMENT)
 350                 info->uses_kill = nir->info.fs.uses_discard;
 351
 352         /* TODO make this more accurate */
 353         info->const_buffers_declared = u_bit_consecutive(0, SI_NUM_CONST_BUFFERS);
 354         info->shader_buffers_declared = u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS);
 355
 356         func = (struct nir_function *)exec_list_get_head_const(&nir->functions);
 357         nir_foreach_block(block, func->impl) {
 358                 nir_foreach_instr(instr, block)
 359                         scan_instruction(info, instr);
 360         }
 361 }
 362
 363 /**
 364  * Perform "lowering" operations on the NIR that are run once when the shader
 365  * selector is created.
 366  */
 367 void
 368 si_lower_nir(struct si_shader_selector* sel)
 369 {
 370         /* Adjust the driver location of inputs and outputs. The state tracker
 371          * interprets them as slots, while the ac/nir backend interprets them
 372          * as individual components.
 373          */
 374         nir_foreach_variable(variable, &sel->nir->inputs)
 375                 variable->data.driver_location *= 4;
 376
 377         nir_foreach_variable(variable, &sel->nir->outputs) {
 378                 variable->data.driver_location *= 4;
 379
 380                 if (sel->nir->info.stage == MESA_SHADER_FRAGMENT) {
 381                         if (variable->data.location == FRAG_RESULT_DEPTH)
 382                                 variable->data.driver_location += 2;
 383                         else if (variable->data.location == FRAG_RESULT_STENCIL)
 384                                 variable->data.driver_location += 1;
 385                 }
 386         }
 387
 388         /* Perform lowerings (and optimizations) of code.
 389          *
 390          * Performance considerations aside, we must:
 391          * - lower certain ALU operations
 392          * - ensure constant offsets for texture instructions are folded
 393          *   and copy-propagated
 394          */
 395         NIR_PASS_V(sel->nir, nir_lower_io, nir_var_uniform, type_size,
 396                    (nir_lower_io_options)0);
 397         NIR_PASS_V(sel->nir, nir_lower_uniforms_to_ubo);
 398
 399         NIR_PASS_V(sel->nir, nir_lower_returns);
 400         NIR_PASS_V(sel->nir, nir_lower_vars_to_ssa);
 401         NIR_PASS_V(sel->nir, nir_lower_alu_to_scalar);
 402         NIR_PASS_V(sel->nir, nir_lower_phis_to_scalar);
 403
 404         static const struct nir_lower_tex_options lower_tex_options = {
 405                 .lower_txp = ~0u,
 406         };
 407         NIR_PASS_V(sel->nir, nir_lower_tex, &lower_tex_options);
 408
 409         bool progress;
 410         do {
 411                 progress = false;
 412
 413                 /* (Constant) copy propagation is needed for txf with offsets. */
 414                 NIR_PASS(progress, sel->nir, nir_copy_prop);
 415                 NIR_PASS(progress, sel->nir, nir_opt_remove_phis);
 416                 NIR_PASS(progress, sel->nir, nir_opt_dce);
 417                 if (nir_opt_trivial_continues(sel->nir)) {
 418                         progress = true;
 419                         NIR_PASS(progress, sel->nir, nir_copy_prop);
 420                         NIR_PASS(progress, sel->nir, nir_opt_dce);
 421                 }
 422                 NIR_PASS(progress, sel->nir, nir_opt_if);
 423                 NIR_PASS(progress, sel->nir, nir_opt_dead_cf);
 424                 NIR_PASS(progress, sel->nir, nir_opt_cse);
 425                 NIR_PASS(progress, sel->nir, nir_opt_peephole_select, 8);
 426
 427                 /* Needed for algebraic lowering */
 428                 NIR_PASS(progress, sel->nir, nir_opt_algebraic);
 429                 NIR_PASS(progress, sel->nir, nir_opt_constant_folding);
 430
 431                 NIR_PASS(progress, sel->nir, nir_opt_undef);
 432                 NIR_PASS(progress, sel->nir, nir_opt_conditional_discard);
 433                 if (sel->nir->options->max_unroll_iterations) {
 434                         NIR_PASS(progress, sel->nir, nir_opt_loop_unroll, 0);
 435                 }
 436         } while (progress);
 437 }
 438
 439 static void declare_nir_input_vs(struct si_shader_context *ctx,
 440                                  struct nir_variable *variable, unsigned rel,
 441                                  LLVMValueRef out[4])
 442 {
 443         si_llvm_load_input_vs(ctx, variable->data.driver_location / 4 + rel, out);
 444 }
 445
 446 static void declare_nir_input_fs(struct si_shader_context *ctx,
 447                                  struct nir_variable *variable, unsigned rel,
 448                                  unsigned *fs_attr_idx,
 449                                  LLVMValueRef out[4])
 450 {
 451         unsigned slot = variable->data.location + rel;
 452
 453         assert(variable->data.location >= VARYING_SLOT_VAR0 || rel == 0);
 454
 455         if (slot == VARYING_SLOT_POS) {
 456                 out[0] = LLVMGetParam(ctx->main_fn, SI_PARAM_POS_X_FLOAT);
 457                 out[1] = LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Y_FLOAT);
 458                 out[2] = LLVMGetParam(ctx->main_fn, SI_PARAM_POS_Z_FLOAT);
 459                 out[3] = ac_build_fdiv(&ctx->ac, ctx->ac.f32_1,
 460                                 LLVMGetParam(ctx->main_fn, SI_PARAM_POS_W_FLOAT));
 461                 return;
 462         }
 463
 464         si_llvm_load_input_fs(ctx, *fs_attr_idx, out);
 465         (*fs_attr_idx)++;
 466 }
 467
 468 static LLVMValueRef
 469 si_nir_load_sampler_desc(struct ac_shader_abi *abi,
 470                          unsigned descriptor_set, unsigned base_index,
 471                          unsigned constant_index, LLVMValueRef dynamic_index,
 472                          enum ac_descriptor_type desc_type, bool image,
 473                          bool write)
 474 {
 475         struct si_shader_context *ctx = si_shader_context_from_abi(abi);
 476         LLVMBuilderRef builder = ctx->ac.builder;
 477         LLVMValueRef list = LLVMGetParam(ctx->main_fn, ctx->param_samplers_and_images);
 478         LLVMValueRef index = dynamic_index;
 479
 480         assert(!descriptor_set);
 481
 482         if (!index)
 483                 index = ctx->ac.i32_0;
 484
 485         index = LLVMBuildAdd(builder, index,
 486                              LLVMConstInt(ctx->ac.i32, base_index + constant_index, false),
 487                              "");
 488
 489         if (image) {
 490                 assert(desc_type == AC_DESC_IMAGE || desc_type == AC_DESC_BUFFER);
 491                 assert(base_index + constant_index < ctx->num_images);
 492
 493                 if (dynamic_index)
 494                         index = si_llvm_bound_index(ctx, index, ctx->num_images);
 495
 496                 index = LLVMBuildSub(ctx->gallivm.builder,
 497                                      LLVMConstInt(ctx->i32, SI_NUM_IMAGES - 1, 0),
 498                                      index, "");
 499
 500                 /* TODO: be smarter about when we use dcc_off */
 501                 return si_load_image_desc(ctx, list, index, desc_type, write);
 502         }
 503
 504         assert(base_index + constant_index < ctx->num_samplers);
 505
 506         if (dynamic_index)
 507                 index = si_llvm_bound_index(ctx, index, ctx->num_samplers);
 508
 509         index = LLVMBuildAdd(ctx->gallivm.builder, index,
 510                              LLVMConstInt(ctx->i32, SI_NUM_IMAGES / 2, 0), "");
 511
 512         return si_load_sampler_desc(ctx, list, index, desc_type);
 513 }
 514
 515 bool si_nir_build_llvm(struct si_shader_context *ctx, struct nir_shader *nir)
 516 {
 517         struct tgsi_shader_info *info = &ctx->shader->selector->info;
 518
 519         unsigned fs_attr_idx = 0;
 520         nir_foreach_variable(variable, &nir->inputs) {
 521                 unsigned attrib_count = glsl_count_attribute_slots(variable->type,
 522                                                                    nir->info.stage == MESA_SHADER_VERTEX);
 523                 unsigned input_idx = variable->data.driver_location;
 524
 525                 for (unsigned i = 0; i < attrib_count; ++i) {
 526                         LLVMValueRef data[4];
 527
 528                         if (nir->info.stage == MESA_SHADER_VERTEX)
 529                                 declare_nir_input_vs(ctx, variable, i, data);
 530                         else if (nir->info.stage == MESA_SHADER_FRAGMENT)
 531                                 declare_nir_input_fs(ctx, variable, i, &fs_attr_idx, data);
 532
 533                         for (unsigned chan = 0; chan < 4; chan++) {
 534                                 ctx->inputs[input_idx + chan] =
 535                                         LLVMBuildBitCast(ctx->ac.builder, data[chan], ctx->ac.i32, "");
 536                         }
 537                 }
 538         }
 539
 540         ctx->abi.inputs = &ctx->inputs[0];
 541         ctx->abi.load_sampler_desc = si_nir_load_sampler_desc;
 542         ctx->abi.clamp_shadow_reference = true;
 543
 544         ctx->num_samplers = util_last_bit(info->samplers_declared);
 545         ctx->num_images = util_last_bit(info->images_declared);
 546
 547         ac_nir_translate(&ctx->ac, &ctx->abi, nir, NULL);
 548
 549         return true;
 550 }