src/gallium/drivers/lima/lima_program.c

   1 /*
   2  * Copyright (c) 2017-2019 Lima Project
   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, sub license,
   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
  12  * next paragraph) shall be included in all copies or substantial portions
  13  * of the 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 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 #include "util/u_memory.h"
  26 #include "util/ralloc.h"
  27 #include "util/u_debug.h"
  28
  29 #include "tgsi/tgsi_dump.h"
  30 #include "compiler/nir/nir.h"
  31 #include "nir/tgsi_to_nir.h"
  32
  33 #include "pipe/p_state.h"
  34
  35 #include "lima_screen.h"
  36 #include "lima_context.h"
  37 #include "lima_job.h"
  38 #include "lima_program.h"
  39 #include "lima_bo.h"
  40 #include "lima_format.h"
  41
  42 #include "ir/lima_ir.h"
  43
  44 static const nir_shader_compiler_options vs_nir_options = {
  45    .lower_ffma = true,
  46    .lower_fpow = true,
  47    .lower_ffract = true,
  48    .lower_fdiv = true,
  49    .lower_fmod = true,
  50    .lower_fsqrt = true,
  51    .lower_sub = true,
  52    .lower_flrp32 = true,
  53    .lower_flrp64 = true,
  54    /* could be implemented by clamp */
  55    .lower_fsat = true,
  56    .lower_bitops = true,
  57    .lower_rotate = true,
  58    .lower_sincos = true,
  59    .lower_fceil = true,
  60 };
  61
  62 static const nir_shader_compiler_options fs_nir_options = {
  63    .lower_ffma = true,
  64    .lower_fpow = true,
  65    .lower_fdiv = true,
  66    .lower_fmod = true,
  67    .lower_sub = true,
  68    .lower_flrp32 = true,
  69    .lower_flrp64 = true,
  70    .lower_fsign = true,
  71    .lower_rotate = true,
  72    .lower_fdot = true,
  73    .lower_fdph = true,
  74    .lower_bitops = true,
  75    .lower_vector_cmp = true,
  76 };
  77
  78 const void *
  79 lima_program_get_compiler_options(enum pipe_shader_type shader)
  80 {
  81    switch (shader) {
  82    case PIPE_SHADER_VERTEX:
  83       return &vs_nir_options;
  84    case PIPE_SHADER_FRAGMENT:
  85       return &fs_nir_options;
  86    default:
  87       return NULL;
  88    }
  89 }
  90
  91 static int
  92 type_size(const struct glsl_type *type, bool bindless)
  93 {
  94    return glsl_count_attribute_slots(type, false);
  95 }
  96
  97 void
  98 lima_program_optimize_vs_nir(struct nir_shader *s)
  99 {
 100    bool progress;
 101
 102    NIR_PASS_V(s, nir_lower_viewport_transform);
 103    NIR_PASS_V(s, nir_lower_point_size, 1.0f, 100.0f);
 104    NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size, 0);
 105    NIR_PASS_V(s, nir_lower_load_const_to_scalar);
 106    NIR_PASS_V(s, lima_nir_lower_uniform_to_scalar);
 107    NIR_PASS_V(s, nir_lower_io_to_scalar,
 108               nir_var_shader_in|nir_var_shader_out);
 109
 110    do {
 111       progress = false;
 112
 113       NIR_PASS_V(s, nir_lower_vars_to_ssa);
 114       NIR_PASS(progress, s, nir_lower_alu_to_scalar, NULL, NULL);
 115       NIR_PASS(progress, s, nir_lower_phis_to_scalar);
 116       NIR_PASS(progress, s, nir_copy_prop);
 117       NIR_PASS(progress, s, nir_opt_remove_phis);
 118       NIR_PASS(progress, s, nir_opt_dce);
 119       NIR_PASS(progress, s, nir_opt_dead_cf);
 120       NIR_PASS(progress, s, nir_opt_cse);
 121       NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
 122       NIR_PASS(progress, s, nir_opt_algebraic);
 123       NIR_PASS(progress, s, lima_nir_lower_ftrunc);
 124       NIR_PASS(progress, s, nir_opt_constant_folding);
 125       NIR_PASS(progress, s, nir_opt_undef);
 126       NIR_PASS(progress, s, nir_opt_loop_unroll,
 127                nir_var_shader_in |
 128                nir_var_shader_out |
 129                nir_var_function_temp);
 130    } while (progress);
 131
 132    NIR_PASS_V(s, nir_lower_int_to_float);
 133    /* int_to_float pass generates ftrunc, so lower it */
 134    NIR_PASS(progress, s, lima_nir_lower_ftrunc);
 135    NIR_PASS_V(s, nir_lower_bool_to_float);
 136
 137    NIR_PASS_V(s, nir_copy_prop);
 138    NIR_PASS_V(s, nir_opt_dce);
 139    NIR_PASS_V(s, nir_lower_locals_to_regs);
 140    NIR_PASS_V(s, nir_convert_from_ssa, true);
 141    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
 142    nir_sweep(s);
 143 }
 144
 145 static bool
 146 lima_alu_to_scalar_filter_cb(const nir_instr *instr, const void *data)
 147 {
 148    if (instr->type != nir_instr_type_alu)
 149       return false;
 150
 151    nir_alu_instr *alu = nir_instr_as_alu(instr);
 152    switch (alu->op) {
 153    case nir_op_frcp:
 154    case nir_op_frsq:
 155    case nir_op_flog2:
 156    case nir_op_fexp2:
 157    case nir_op_fsqrt:
 158    case nir_op_fsin:
 159    case nir_op_fcos:
 160       return true;
 161    default:
 162       break;
 163    }
 164
 165    /* nir vec4 fcsel assumes that each component of the condition will be
 166     * used to select the same component from the two options, but Utgard PP
 167     * has only 1 component condition. If all condition components are not the
 168     * same we need to lower it to scalar.
 169     */
 170    switch (alu->op) {
 171    case nir_op_bcsel:
 172    case nir_op_fcsel:
 173       break;
 174    default:
 175       return false;
 176    }
 177
 178    int num_components = nir_dest_num_components(alu->dest.dest);
 179
 180    uint8_t swizzle = alu->src[0].swizzle[0];
 181
 182    for (int i = 1; i < num_components; i++)
 183       if (alu->src[0].swizzle[i] != swizzle)
 184          return true;
 185
 186    return false;
 187 }
 188
 189 void
 190 lima_program_optimize_fs_nir(struct nir_shader *s,
 191                              struct nir_lower_tex_options *tex_options)
 192 {
 193    bool progress;
 194
 195    NIR_PASS_V(s, nir_lower_fragcoord_wtrans);
 196    NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size, 0);
 197    NIR_PASS_V(s, nir_lower_regs_to_ssa);
 198    NIR_PASS_V(s, nir_lower_tex, tex_options);
 199
 200    do {
 201       progress = false;
 202       NIR_PASS(progress, s, nir_opt_vectorize);
 203    } while (progress);
 204
 205    do {
 206       progress = false;
 207
 208       NIR_PASS_V(s, nir_lower_vars_to_ssa);
 209       NIR_PASS(progress, s, nir_lower_alu_to_scalar, lima_alu_to_scalar_filter_cb, NULL);
 210       NIR_PASS(progress, s, nir_copy_prop);
 211       NIR_PASS(progress, s, nir_opt_remove_phis);
 212       NIR_PASS(progress, s, nir_opt_dce);
 213       NIR_PASS(progress, s, nir_opt_dead_cf);
 214       NIR_PASS(progress, s, nir_opt_cse);
 215       NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
 216       NIR_PASS(progress, s, nir_opt_algebraic);
 217       NIR_PASS(progress, s, nir_opt_constant_folding);
 218       NIR_PASS(progress, s, nir_opt_undef);
 219       NIR_PASS(progress, s, nir_opt_loop_unroll,
 220                nir_var_shader_in |
 221                nir_var_shader_out |
 222                nir_var_function_temp);
 223       NIR_PASS(progress, s, lima_nir_split_load_input);
 224    } while (progress);
 225
 226    NIR_PASS_V(s, nir_lower_int_to_float);
 227    NIR_PASS_V(s, nir_lower_bool_to_float);
 228
 229    /* Some ops must be lowered after being converted from int ops,
 230     * so re-run nir_opt_algebraic after int lowering. */
 231    do {
 232       progress = false;
 233       NIR_PASS(progress, s, nir_opt_algebraic);
 234    } while (progress);
 235
 236    /* Must be run after optimization loop */
 237    NIR_PASS_V(s, lima_nir_scale_trig);
 238
 239    /* Lower modifiers */
 240    NIR_PASS_V(s, nir_lower_to_source_mods, nir_lower_all_source_mods);
 241    NIR_PASS_V(s, nir_copy_prop);
 242    NIR_PASS_V(s, nir_opt_dce);
 243
 244    NIR_PASS_V(s, nir_lower_locals_to_regs);
 245    NIR_PASS_V(s, nir_convert_from_ssa, true);
 246    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
 247
 248    NIR_PASS_V(s, nir_move_vec_src_uses_to_dest);
 249    NIR_PASS_V(s, nir_lower_vec_to_movs);
 250
 251    NIR_PASS_V(s, lima_nir_duplicate_load_uniforms);
 252    NIR_PASS_V(s, lima_nir_duplicate_load_inputs);
 253
 254    nir_sweep(s);
 255 }
 256
 257 static bool
 258 lima_fs_compile_shader(struct lima_context *ctx,
 259                        struct lima_fs_shader_state *fs,
 260                        struct nir_lower_tex_options *tex_options)
 261 {
 262    struct lima_screen *screen = lima_screen(ctx->base.screen);
 263    nir_shader *nir = nir_shader_clone(fs, fs->base.ir.nir);
 264
 265    lima_program_optimize_fs_nir(nir, tex_options);
 266
 267    if (lima_debug & LIMA_DEBUG_PP)
 268       nir_print_shader(nir, stdout);
 269
 270    if (!ppir_compile_nir(fs, nir, screen->pp_ra, &ctx->debug)) {
 271       ralloc_free(nir);
 272       return false;
 273    }
 274
 275    fs->uses_discard = nir->info.fs.uses_discard;
 276    ralloc_free(nir);
 277
 278    return true;
 279 }
 280
 281 static void *
 282 lima_create_fs_state(struct pipe_context *pctx,
 283                      const struct pipe_shader_state *cso)
 284 {
 285    struct lima_context *ctx = lima_context(pctx);
 286    struct lima_fs_shader_state *so = rzalloc(NULL, struct lima_fs_shader_state);
 287
 288    if (!so)
 289       return NULL;
 290
 291    nir_shader *nir;
 292    if (cso->type == PIPE_SHADER_IR_NIR)
 293       /* The backend takes ownership of the NIR shader on state
 294        * creation.
 295        */
 296       nir = cso->ir.nir;
 297    else {
 298       assert(cso->type == PIPE_SHADER_IR_TGSI);
 299
 300       nir = tgsi_to_nir(cso->tokens, pctx->screen);
 301    }
 302
 303    so->base.type = PIPE_SHADER_IR_NIR;
 304    so->base.ir.nir = nir;
 305
 306    uint8_t identity[4] = { PIPE_SWIZZLE_X,
 307                            PIPE_SWIZZLE_Y,
 308                            PIPE_SWIZZLE_Z,
 309                            PIPE_SWIZZLE_W };
 310
 311    struct nir_lower_tex_options tex_options = {
 312       .lower_txp = ~0u,
 313       .swizzle_result = 0,
 314    };
 315
 316    /* Initialize with identity swizzles. That should suffice for most shaders  */
 317    for (int i = 0; i < PIPE_MAX_SAMPLERS; i++)
 318       memcpy(so->swizzles[i], identity, 4);
 319
 320    if (!lima_fs_compile_shader(ctx, so, &tex_options)) {
 321       ralloc_free(so);
 322       return NULL;
 323    }
 324
 325    return so;
 326 }
 327
 328 static void
 329 lima_bind_fs_state(struct pipe_context *pctx, void *hwcso)
 330 {
 331    struct lima_context *ctx = lima_context(pctx);
 332
 333    ctx->fs = hwcso;
 334    ctx->dirty |= LIMA_CONTEXT_DIRTY_SHADER_FRAG;
 335 }
 336
 337 static void
 338 lima_delete_fs_state(struct pipe_context *pctx, void *hwcso)
 339 {
 340    struct lima_fs_shader_state *so = hwcso;
 341
 342    if (so->bo)
 343       lima_bo_unreference(so->bo);
 344
 345    ralloc_free(so->base.ir.nir);
 346    ralloc_free(so);
 347 }
 348
 349 bool
 350 lima_update_vs_state(struct lima_context *ctx)
 351 {
 352    struct lima_vs_shader_state *vs = ctx->vs;
 353    if (!vs->bo) {
 354       struct lima_screen *screen = lima_screen(ctx->base.screen);
 355       vs->bo = lima_bo_create(screen, vs->shader_size, 0);
 356       if (!vs->bo) {
 357          fprintf(stderr, "lima: create vs shader bo fail\n");
 358          return false;
 359       }
 360
 361       memcpy(lima_bo_map(vs->bo), vs->shader, vs->shader_size);
 362       ralloc_free(vs->shader);
 363       vs->shader = NULL;
 364    }
 365
 366    return true;
 367 }
 368
 369 bool
 370 lima_update_fs_state(struct lima_context *ctx)
 371 {
 372    struct lima_fs_shader_state *fs = ctx->fs;
 373    struct lima_texture_stateobj *lima_tex = &ctx->tex_stateobj;
 374    struct nir_lower_tex_options tex_options = {
 375       .lower_txp = ~0u,
 376       .swizzle_result = 0,
 377    };
 378    bool needs_recompile = false;
 379
 380    /* Check if texture formats has changed since last compilation.
 381     * If it has we need to recompile shader.
 382     */
 383    if (((ctx->dirty & LIMA_CONTEXT_DIRTY_TEXTURES) &&
 384        lima_tex->num_samplers &&
 385        lima_tex->num_textures)) {
 386       uint8_t identity[4] = { PIPE_SWIZZLE_X,
 387                               PIPE_SWIZZLE_Y,
 388                               PIPE_SWIZZLE_Z,
 389                               PIPE_SWIZZLE_W };
 390       for (int i = 0; i < lima_tex->num_samplers; i++) {
 391          struct lima_sampler_view *texture = lima_sampler_view(lima_tex->textures[i]);
 392          struct pipe_resource *prsc = texture->base.texture;
 393          const uint8_t *swizzle = lima_format_get_texel_swizzle(prsc->format);
 394          if (memcmp(fs->swizzles[i], swizzle, 4)) {
 395             needs_recompile = true;
 396             memcpy(fs->swizzles[i], swizzle, 4);
 397          }
 398
 399          for (int j = 0; j < 4; j++)
 400             tex_options.swizzles[i][j] = swizzle[j];
 401
 402          if (memcmp(swizzle, identity, 4))
 403             tex_options.swizzle_result |= (1 << i);
 404       }
 405
 406       /* Fill rest with identity swizzle */
 407       for (int i = lima_tex->num_samplers; i < PIPE_MAX_SAMPLERS; i++)
 408          memcpy(fs->swizzles[i], identity, 4);
 409    }
 410
 411    if (needs_recompile) {
 412       if (fs->bo) {
 413          lima_bo_unreference(fs->bo);
 414          fs->bo = NULL;
 415       }
 416
 417       if (!lima_fs_compile_shader(ctx, fs, &tex_options))
 418          return false;
 419    }
 420
 421    if (!fs->bo) {
 422       struct lima_screen *screen = lima_screen(ctx->base.screen);
 423       fs->bo = lima_bo_create(screen, fs->shader_size, 0);
 424       if (!fs->bo) {
 425          fprintf(stderr, "lima: create fs shader bo fail\n");
 426          return false;
 427       }
 428
 429       memcpy(lima_bo_map(fs->bo), fs->shader, fs->shader_size);
 430       ralloc_free(fs->shader);
 431       fs->shader = NULL;
 432    }
 433
 434    struct lima_job *job = lima_job_get(ctx);
 435    job->pp_max_stack_size = MAX2(job->pp_max_stack_size, ctx->fs->stack_size);
 436
 437    return true;
 438 }
 439
 440 static void *
 441 lima_create_vs_state(struct pipe_context *pctx,
 442                      const struct pipe_shader_state *cso)
 443 {
 444    struct lima_context *ctx = lima_context(pctx);
 445    struct lima_vs_shader_state *so = rzalloc(NULL, struct lima_vs_shader_state);
 446
 447    if (!so)
 448       return NULL;
 449
 450    nir_shader *nir;
 451    if (cso->type == PIPE_SHADER_IR_NIR)
 452       nir = cso->ir.nir;
 453    else {
 454       assert(cso->type == PIPE_SHADER_IR_TGSI);
 455
 456       nir = tgsi_to_nir(cso->tokens, pctx->screen);
 457    }
 458
 459    lima_program_optimize_vs_nir(nir);
 460
 461    if (lima_debug & LIMA_DEBUG_GP)
 462       nir_print_shader(nir, stdout);
 463
 464    if (!gpir_compile_nir(so, nir, &ctx->debug)) {
 465       ralloc_free(so);
 466       return NULL;
 467    }
 468
 469    ralloc_free(nir);
 470
 471    return so;
 472 }
 473
 474 static void
 475 lima_bind_vs_state(struct pipe_context *pctx, void *hwcso)
 476 {
 477    struct lima_context *ctx = lima_context(pctx);
 478
 479    ctx->vs = hwcso;
 480    ctx->dirty |= LIMA_CONTEXT_DIRTY_SHADER_VERT;
 481 }
 482
 483 static void
 484 lima_delete_vs_state(struct pipe_context *pctx, void *hwcso)
 485 {
 486    struct lima_vs_shader_state *so = hwcso;
 487
 488    if (so->bo)
 489       lima_bo_unreference(so->bo);
 490
 491    ralloc_free(so);
 492 }
 493
 494 void
 495 lima_program_init(struct lima_context *ctx)
 496 {
 497    ctx->base.create_fs_state = lima_create_fs_state;
 498    ctx->base.bind_fs_state = lima_bind_fs_state;
 499    ctx->base.delete_fs_state = lima_delete_fs_state;
 500
 501    ctx->base.create_vs_state = lima_create_vs_state;
 502    ctx->base.bind_vs_state = lima_bind_vs_state;
 503    ctx->base.delete_vs_state = lima_delete_vs_state;
 504 }