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_program.h"
  38 #include "lima_bo.h"
  39 #include "ir/lima_ir.h"
  40
  41 static const nir_shader_compiler_options vs_nir_options = {
  42    .lower_ffma = true,
  43    .lower_fpow = true,
  44    .lower_ffract = true,
  45    .lower_fdiv = true,
  46    .lower_fmod = true,
  47    .lower_fsqrt = true,
  48    .lower_sub = true,
  49    .lower_flrp32 = true,
  50    .lower_flrp64 = true,
  51    .lower_ftrunc = true,
  52    /* could be implemented by clamp */
  53    .lower_fsat = true,
  54    .lower_bitops = true,
  55    .lower_rotate = true,
  56    .lower_sincos = true,
  57    .lower_fceil = true,
  58 };
  59
  60 static const nir_shader_compiler_options fs_nir_options = {
  61    .lower_ffma = true,
  62    .lower_fpow = true,
  63    .lower_fdiv = true,
  64    .lower_fmod = true,
  65    .lower_sub = true,
  66    .lower_flrp32 = true,
  67    .lower_flrp64 = true,
  68    .lower_fsign = true,
  69    .lower_rotate = true,
  70    .lower_fdot = true,
  71    .lower_fdph = true,
  72    .lower_bitops = true,
  73    .lower_vector_cmp = true,
  74 };
  75
  76 static const struct nir_lower_tex_options tex_options = {
  77    .lower_txp = ~0u,
  78 };
  79
  80 const void *
  81 lima_program_get_compiler_options(enum pipe_shader_type shader)
  82 {
  83    switch (shader) {
  84    case PIPE_SHADER_VERTEX:
  85       return &vs_nir_options;
  86    case PIPE_SHADER_FRAGMENT:
  87       return &fs_nir_options;
  88    default:
  89       return NULL;
  90    }
  91 }
  92
  93 static int
  94 type_size(const struct glsl_type *type, bool bindless)
  95 {
  96    return glsl_count_attribute_slots(type, false);
  97 }
  98
  99 void
 100 lima_program_optimize_vs_nir(struct nir_shader *s)
 101 {
 102    bool progress;
 103
 104    NIR_PASS_V(s, nir_lower_viewport_transform);
 105    NIR_PASS_V(s, nir_lower_point_size, 1.0f, 100.0f);
 106    NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size, 0);
 107    NIR_PASS_V(s, nir_lower_load_const_to_scalar);
 108    NIR_PASS_V(s, lima_nir_lower_uniform_to_scalar);
 109    NIR_PASS_V(s, nir_lower_io_to_scalar,
 110               nir_var_shader_in|nir_var_shader_out);
 111
 112    do {
 113       progress = false;
 114
 115       NIR_PASS_V(s, nir_lower_vars_to_ssa);
 116       NIR_PASS(progress, s, nir_lower_alu_to_scalar, NULL, NULL);
 117       NIR_PASS(progress, s, nir_lower_phis_to_scalar);
 118       NIR_PASS(progress, s, nir_copy_prop);
 119       NIR_PASS(progress, s, nir_opt_remove_phis);
 120       NIR_PASS(progress, s, nir_opt_dce);
 121       NIR_PASS(progress, s, nir_opt_dead_cf);
 122       NIR_PASS(progress, s, nir_opt_cse);
 123       NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
 124       NIR_PASS(progress, s, nir_opt_algebraic);
 125       NIR_PASS(progress, s, nir_opt_constant_folding);
 126       NIR_PASS(progress, s, nir_opt_undef);
 127       NIR_PASS(progress, s, nir_opt_loop_unroll,
 128                nir_var_shader_in |
 129                nir_var_shader_out |
 130                nir_var_function_temp);
 131    } while (progress);
 132
 133    NIR_PASS_V(s, nir_lower_int_to_float);
 134    /* Run opt_algebraic between int_to_float and bool_to_float because
 135     * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
 136     */
 137    do {
 138       progress = false;
 139       NIR_PASS(progress, s, nir_opt_algebraic);
 140    } while (progress);
 141
 142    NIR_PASS_V(s, nir_lower_bool_to_float);
 143
 144    NIR_PASS_V(s, nir_copy_prop);
 145    NIR_PASS_V(s, nir_opt_dce);
 146    NIR_PASS_V(s, nir_lower_locals_to_regs);
 147    NIR_PASS_V(s, nir_convert_from_ssa, true);
 148    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
 149    nir_sweep(s);
 150 }
 151
 152 static bool
 153 lima_alu_to_scalar_filter_cb(const nir_instr *instr, const void *data)
 154 {
 155    if (instr->type != nir_instr_type_alu)
 156       return false;
 157
 158    nir_alu_instr *alu = nir_instr_as_alu(instr);
 159    switch (alu->op) {
 160    case nir_op_frcp:
 161    case nir_op_frsq:
 162    case nir_op_flog2:
 163    case nir_op_fexp2:
 164    case nir_op_fsqrt:
 165    case nir_op_fsin:
 166    case nir_op_fcos:
 167       return true;
 168    default:
 169       break;
 170    }
 171
 172    /* nir vec4 fcsel assumes that each component of the condition will be
 173     * used to select the same component from the two options, but Utgard PP
 174     * has only 1 component condition. If all condition components are not the
 175     * same we need to lower it to scalar.
 176     */
 177    switch (alu->op) {
 178    case nir_op_bcsel:
 179    case nir_op_fcsel:
 180       break;
 181    default:
 182       return false;
 183    }
 184
 185    int num_components = nir_dest_num_components(alu->dest.dest);
 186
 187    uint8_t swizzle = alu->src[0].swizzle[0];
 188
 189    for (int i = 1; i < num_components; i++)
 190       if (alu->src[0].swizzle[i] != swizzle)
 191          return true;
 192
 193    return false;
 194 }
 195
 196 void
 197 lima_program_optimize_fs_nir(struct nir_shader *s)
 198 {
 199    bool progress;
 200
 201    NIR_PASS_V(s, nir_lower_fragcoord_wtrans);
 202    NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size, 0);
 203    NIR_PASS_V(s, nir_lower_regs_to_ssa);
 204    NIR_PASS_V(s, nir_lower_tex, &tex_options);
 205
 206    do {
 207       progress = false;
 208       NIR_PASS(progress, s, nir_opt_vectorize);
 209    } while (progress);
 210
 211    do {
 212       progress = false;
 213
 214       NIR_PASS_V(s, nir_lower_vars_to_ssa);
 215       NIR_PASS(progress, s, nir_lower_alu_to_scalar, lima_alu_to_scalar_filter_cb, NULL);
 216       NIR_PASS(progress, s, nir_copy_prop);
 217       NIR_PASS(progress, s, nir_opt_remove_phis);
 218       NIR_PASS(progress, s, nir_opt_dce);
 219       NIR_PASS(progress, s, nir_opt_dead_cf);
 220       NIR_PASS(progress, s, nir_opt_cse);
 221       NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
 222       NIR_PASS(progress, s, nir_opt_algebraic);
 223       NIR_PASS(progress, s, nir_opt_constant_folding);
 224       NIR_PASS(progress, s, nir_opt_undef);
 225       NIR_PASS(progress, s, nir_opt_loop_unroll,
 226                nir_var_shader_in |
 227                nir_var_shader_out |
 228                nir_var_function_temp);
 229       NIR_PASS(progress, s, lima_nir_split_load_input);
 230    } while (progress);
 231
 232    NIR_PASS_V(s, nir_lower_int_to_float);
 233    NIR_PASS_V(s, nir_lower_bool_to_float);
 234
 235    /* Some ops must be lowered after being converted from int ops,
 236     * so re-run nir_opt_algebraic after int lowering. */
 237    do {
 238       progress = false;
 239       NIR_PASS(progress, s, nir_opt_algebraic);
 240    } while (progress);
 241
 242    /* Must be run after optimization loop */
 243    NIR_PASS_V(s, lima_nir_scale_trig);
 244
 245    /* Lower modifiers */
 246    NIR_PASS_V(s, nir_lower_to_source_mods, nir_lower_all_source_mods);
 247    NIR_PASS_V(s, nir_copy_prop);
 248    NIR_PASS_V(s, nir_opt_dce);
 249
 250    NIR_PASS_V(s, nir_lower_locals_to_regs);
 251    NIR_PASS_V(s, nir_convert_from_ssa, true);
 252    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
 253
 254    NIR_PASS_V(s, nir_move_vec_src_uses_to_dest);
 255    NIR_PASS_V(s, nir_lower_vec_to_movs);
 256
 257    nir_sweep(s);
 258 }
 259
 260 static void *
 261 lima_create_fs_state(struct pipe_context *pctx,
 262                      const struct pipe_shader_state *cso)
 263 {
 264    struct lima_context *ctx = lima_context(pctx);
 265    struct lima_screen *screen = lima_screen(pctx->screen);
 266    struct lima_fs_shader_state *so = rzalloc(NULL, struct lima_fs_shader_state);
 267
 268    if (!so)
 269       return NULL;
 270
 271    nir_shader *nir;
 272    if (cso->type == PIPE_SHADER_IR_NIR)
 273       nir = cso->ir.nir;
 274    else {
 275       assert(cso->type == PIPE_SHADER_IR_TGSI);
 276
 277       nir = tgsi_to_nir(cso->tokens, pctx->screen);
 278    }
 279
 280    lima_program_optimize_fs_nir(nir);
 281
 282    if (lima_debug & LIMA_DEBUG_PP)
 283       nir_print_shader(nir, stdout);
 284
 285    if (!ppir_compile_nir(so, nir, screen->pp_ra, &ctx->debug)) {
 286       ralloc_free(so);
 287       return NULL;
 288    }
 289
 290    so->uses_discard = nir->info.fs.uses_discard;
 291
 292    return so;
 293 }
 294
 295 static void
 296 lima_bind_fs_state(struct pipe_context *pctx, void *hwcso)
 297 {
 298    struct lima_context *ctx = lima_context(pctx);
 299
 300    ctx->fs = hwcso;
 301    ctx->dirty |= LIMA_CONTEXT_DIRTY_SHADER_FRAG;
 302 }
 303
 304 static void
 305 lima_delete_fs_state(struct pipe_context *pctx, void *hwcso)
 306 {
 307    struct lima_fs_shader_state *so = hwcso;
 308
 309    if (so->bo)
 310       lima_bo_unreference(so->bo);
 311
 312    ralloc_free(so);
 313 }
 314
 315 bool
 316 lima_update_vs_state(struct lima_context *ctx)
 317 {
 318    struct lima_vs_shader_state *vs = ctx->vs;
 319    if (!vs->bo) {
 320       struct lima_screen *screen = lima_screen(ctx->base.screen);
 321       vs->bo = lima_bo_create(screen, vs->shader_size, 0);
 322       if (!vs->bo) {
 323          fprintf(stderr, "lima: create vs shader bo fail\n");
 324          return false;
 325       }
 326
 327       memcpy(lima_bo_map(vs->bo), vs->shader, vs->shader_size);
 328       ralloc_free(vs->shader);
 329       vs->shader = NULL;
 330    }
 331
 332    return true;
 333 }
 334
 335 bool
 336 lima_update_fs_state(struct lima_context *ctx)
 337 {
 338    struct lima_fs_shader_state *fs = ctx->fs;
 339    if (!fs->bo) {
 340       struct lima_screen *screen = lima_screen(ctx->base.screen);
 341       fs->bo = lima_bo_create(screen, fs->shader_size, 0);
 342       if (!fs->bo) {
 343          fprintf(stderr, "lima: create fs shader bo fail\n");
 344          return false;
 345       }
 346
 347       memcpy(lima_bo_map(fs->bo), fs->shader, fs->shader_size);
 348       ralloc_free(fs->shader);
 349       fs->shader = NULL;
 350    }
 351
 352    ctx->pp_max_stack_size = MAX2(ctx->pp_max_stack_size, ctx->fs->stack_size);
 353
 354    return true;
 355 }
 356
 357 static void *
 358 lima_create_vs_state(struct pipe_context *pctx,
 359                      const struct pipe_shader_state *cso)
 360 {
 361    struct lima_context *ctx = lima_context(pctx);
 362    struct lima_vs_shader_state *so = rzalloc(NULL, struct lima_vs_shader_state);
 363
 364    if (!so)
 365       return NULL;
 366
 367    nir_shader *nir;
 368    if (cso->type == PIPE_SHADER_IR_NIR)
 369       nir = cso->ir.nir;
 370    else {
 371       assert(cso->type == PIPE_SHADER_IR_TGSI);
 372
 373       nir = tgsi_to_nir(cso->tokens, pctx->screen);
 374    }
 375
 376    lima_program_optimize_vs_nir(nir);
 377
 378    if (lima_debug & LIMA_DEBUG_GP)
 379       nir_print_shader(nir, stdout);
 380
 381    if (!gpir_compile_nir(so, nir, &ctx->debug)) {
 382       ralloc_free(so);
 383       return NULL;
 384    }
 385
 386    ralloc_free(nir);
 387
 388    return so;
 389 }
 390
 391 static void
 392 lima_bind_vs_state(struct pipe_context *pctx, void *hwcso)
 393 {
 394    struct lima_context *ctx = lima_context(pctx);
 395
 396    ctx->vs = hwcso;
 397    ctx->dirty |= LIMA_CONTEXT_DIRTY_SHADER_VERT;
 398 }
 399
 400 static void
 401 lima_delete_vs_state(struct pipe_context *pctx, void *hwcso)
 402 {
 403    struct lima_vs_shader_state *so = hwcso;
 404
 405    if (so->bo)
 406       lima_bo_unreference(so->bo);
 407
 408    ralloc_free(so);
 409 }
 410
 411 void
 412 lima_program_init(struct lima_context *ctx)
 413 {
 414    ctx->base.create_fs_state = lima_create_fs_state;
 415    ctx->base.bind_fs_state = lima_bind_fs_state;
 416    ctx->base.delete_fs_state = lima_delete_fs_state;
 417
 418    ctx->base.create_vs_state = lima_create_vs_state;
 419    ctx->base.bind_vs_state = lima_bind_vs_state;
 420    ctx->base.delete_vs_state = lima_delete_vs_state;
 421 }