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 "ir/lima_ir.h"
  41
  42 static const nir_shader_compiler_options vs_nir_options = {
  43    .lower_ffma = true,
  44    .lower_fpow = true,
  45    .lower_ffract = true,
  46    .lower_fdiv = true,
  47    .lower_fmod = true,
  48    .lower_fsqrt = true,
  49    .lower_sub = true,
  50    .lower_flrp32 = true,
  51    .lower_flrp64 = true,
  52    .lower_ftrunc = true,
  53    /* could be implemented by clamp */
  54    .lower_fsat = true,
  55    .lower_bitops = true,
  56    .lower_rotate = true,
  57    .lower_sincos = true,
  58    .lower_fceil = true,
  59 };
  60
  61 static const nir_shader_compiler_options fs_nir_options = {
  62    .lower_ffma = true,
  63    .lower_fpow = true,
  64    .lower_fdiv = true,
  65    .lower_fmod = true,
  66    .lower_sub = true,
  67    .lower_flrp32 = true,
  68    .lower_flrp64 = true,
  69    .lower_fsign = true,
  70    .lower_rotate = true,
  71    .lower_fdot = true,
  72    .lower_fdph = true,
  73    .lower_bitops = true,
  74    .lower_vector_cmp = true,
  75 };
  76
  77 static const struct nir_lower_tex_options tex_options = {
  78    .lower_txp = ~0u,
  79 };
  80
  81 const void *
  82 lima_program_get_compiler_options(enum pipe_shader_type shader)
  83 {
  84    switch (shader) {
  85    case PIPE_SHADER_VERTEX:
  86       return &vs_nir_options;
  87    case PIPE_SHADER_FRAGMENT:
  88       return &fs_nir_options;
  89    default:
  90       return NULL;
  91    }
  92 }
  93
  94 static int
  95 type_size(const struct glsl_type *type, bool bindless)
  96 {
  97    return glsl_count_attribute_slots(type, false);
  98 }
  99
 100 void
 101 lima_program_optimize_vs_nir(struct nir_shader *s)
 102 {
 103    bool progress;
 104
 105    NIR_PASS_V(s, nir_lower_viewport_transform);
 106    NIR_PASS_V(s, nir_lower_point_size, 1.0f, 100.0f);
 107    NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size, 0);
 108    NIR_PASS_V(s, nir_lower_load_const_to_scalar);
 109    NIR_PASS_V(s, lima_nir_lower_uniform_to_scalar);
 110    NIR_PASS_V(s, nir_lower_io_to_scalar,
 111               nir_var_shader_in|nir_var_shader_out);
 112
 113    do {
 114       progress = false;
 115
 116       NIR_PASS_V(s, nir_lower_vars_to_ssa);
 117       NIR_PASS(progress, s, nir_lower_alu_to_scalar, NULL, NULL);
 118       NIR_PASS(progress, s, nir_lower_phis_to_scalar);
 119       NIR_PASS(progress, s, nir_copy_prop);
 120       NIR_PASS(progress, s, nir_opt_remove_phis);
 121       NIR_PASS(progress, s, nir_opt_dce);
 122       NIR_PASS(progress, s, nir_opt_dead_cf);
 123       NIR_PASS(progress, s, nir_opt_cse);
 124       NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
 125       NIR_PASS(progress, s, nir_opt_algebraic);
 126       NIR_PASS(progress, s, nir_opt_constant_folding);
 127       NIR_PASS(progress, s, nir_opt_undef);
 128       NIR_PASS(progress, s, nir_opt_loop_unroll,
 129                nir_var_shader_in |
 130                nir_var_shader_out |
 131                nir_var_function_temp);
 132    } while (progress);
 133
 134    NIR_PASS_V(s, nir_lower_int_to_float);
 135    /* Run opt_algebraic between int_to_float and bool_to_float because
 136     * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
 137     */
 138    do {
 139       progress = false;
 140       NIR_PASS(progress, s, nir_opt_algebraic);
 141    } while (progress);
 142
 143    NIR_PASS_V(s, nir_lower_bool_to_float);
 144
 145    NIR_PASS_V(s, nir_copy_prop);
 146    NIR_PASS_V(s, nir_opt_dce);
 147    NIR_PASS_V(s, nir_lower_locals_to_regs);
 148    NIR_PASS_V(s, nir_convert_from_ssa, true);
 149    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
 150    nir_sweep(s);
 151 }
 152
 153 static bool
 154 lima_alu_to_scalar_filter_cb(const nir_instr *instr, const void *data)
 155 {
 156    if (instr->type != nir_instr_type_alu)
 157       return false;
 158
 159    nir_alu_instr *alu = nir_instr_as_alu(instr);
 160    switch (alu->op) {
 161    case nir_op_frcp:
 162    case nir_op_frsq:
 163    case nir_op_flog2:
 164    case nir_op_fexp2:
 165    case nir_op_fsqrt:
 166    case nir_op_fsin:
 167    case nir_op_fcos:
 168       return true;
 169    default:
 170       break;
 171    }
 172
 173    /* nir vec4 fcsel assumes that each component of the condition will be
 174     * used to select the same component from the two options, but Utgard PP
 175     * has only 1 component condition. If all condition components are not the
 176     * same we need to lower it to scalar.
 177     */
 178    switch (alu->op) {
 179    case nir_op_bcsel:
 180    case nir_op_fcsel:
 181       break;
 182    default:
 183       return false;
 184    }
 185
 186    int num_components = nir_dest_num_components(alu->dest.dest);
 187
 188    uint8_t swizzle = alu->src[0].swizzle[0];
 189
 190    for (int i = 1; i < num_components; i++)
 191       if (alu->src[0].swizzle[i] != swizzle)
 192          return true;
 193
 194    return false;
 195 }
 196
 197 void
 198 lima_program_optimize_fs_nir(struct nir_shader *s)
 199 {
 200    bool progress;
 201
 202    NIR_PASS_V(s, nir_lower_fragcoord_wtrans);
 203    NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size, 0);
 204    NIR_PASS_V(s, nir_lower_regs_to_ssa);
 205    NIR_PASS_V(s, nir_lower_tex, &tex_options);
 206
 207    do {
 208       progress = false;
 209       NIR_PASS(progress, s, nir_opt_vectorize);
 210    } while (progress);
 211
 212    do {
 213       progress = false;
 214
 215       NIR_PASS_V(s, nir_lower_vars_to_ssa);
 216       NIR_PASS(progress, s, nir_lower_alu_to_scalar, lima_alu_to_scalar_filter_cb, NULL);
 217       NIR_PASS(progress, s, nir_copy_prop);
 218       NIR_PASS(progress, s, nir_opt_remove_phis);
 219       NIR_PASS(progress, s, nir_opt_dce);
 220       NIR_PASS(progress, s, nir_opt_dead_cf);
 221       NIR_PASS(progress, s, nir_opt_cse);
 222       NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
 223       NIR_PASS(progress, s, nir_opt_algebraic);
 224       NIR_PASS(progress, s, nir_opt_constant_folding);
 225       NIR_PASS(progress, s, nir_opt_undef);
 226       NIR_PASS(progress, s, nir_opt_loop_unroll,
 227                nir_var_shader_in |
 228                nir_var_shader_out |
 229                nir_var_function_temp);
 230       NIR_PASS(progress, s, lima_nir_split_load_input);
 231    } while (progress);
 232
 233    NIR_PASS_V(s, nir_lower_int_to_float);
 234    NIR_PASS_V(s, nir_lower_bool_to_float);
 235
 236    /* Some ops must be lowered after being converted from int ops,
 237     * so re-run nir_opt_algebraic after int lowering. */
 238    do {
 239       progress = false;
 240       NIR_PASS(progress, s, nir_opt_algebraic);
 241    } while (progress);
 242
 243    /* Must be run after optimization loop */
 244    NIR_PASS_V(s, lima_nir_scale_trig);
 245
 246    /* Lower modifiers */
 247    NIR_PASS_V(s, nir_lower_to_source_mods, nir_lower_all_source_mods);
 248    NIR_PASS_V(s, nir_copy_prop);
 249    NIR_PASS_V(s, nir_opt_dce);
 250
 251    NIR_PASS_V(s, nir_lower_locals_to_regs);
 252    NIR_PASS_V(s, nir_convert_from_ssa, true);
 253    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp);
 254
 255    NIR_PASS_V(s, nir_move_vec_src_uses_to_dest);
 256    NIR_PASS_V(s, nir_lower_vec_to_movs);
 257
 258    nir_sweep(s);
 259 }
 260
 261 static void *
 262 lima_create_fs_state(struct pipe_context *pctx,
 263                      const struct pipe_shader_state *cso)
 264 {
 265    struct lima_context *ctx = lima_context(pctx);
 266    struct lima_screen *screen = lima_screen(pctx->screen);
 267    struct lima_fs_shader_state *so = rzalloc(NULL, struct lima_fs_shader_state);
 268
 269    if (!so)
 270       return NULL;
 271
 272    nir_shader *nir;
 273    if (cso->type == PIPE_SHADER_IR_NIR)
 274       nir = cso->ir.nir;
 275    else {
 276       assert(cso->type == PIPE_SHADER_IR_TGSI);
 277
 278       nir = tgsi_to_nir(cso->tokens, pctx->screen);
 279    }
 280
 281    lima_program_optimize_fs_nir(nir);
 282
 283    if (lima_debug & LIMA_DEBUG_PP)
 284       nir_print_shader(nir, stdout);
 285
 286    if (!ppir_compile_nir(so, nir, screen->pp_ra, &ctx->debug)) {
 287       ralloc_free(so);
 288       return NULL;
 289    }
 290
 291    so->uses_discard = nir->info.fs.uses_discard;
 292
 293    return so;
 294 }
 295
 296 static void
 297 lima_bind_fs_state(struct pipe_context *pctx, void *hwcso)
 298 {
 299    struct lima_context *ctx = lima_context(pctx);
 300
 301    ctx->fs = hwcso;
 302    ctx->dirty |= LIMA_CONTEXT_DIRTY_SHADER_FRAG;
 303 }
 304
 305 static void
 306 lima_delete_fs_state(struct pipe_context *pctx, void *hwcso)
 307 {
 308    struct lima_fs_shader_state *so = hwcso;
 309
 310    if (so->bo)
 311       lima_bo_unreference(so->bo);
 312
 313    ralloc_free(so);
 314 }
 315
 316 bool
 317 lima_update_vs_state(struct lima_context *ctx)
 318 {
 319    struct lima_vs_shader_state *vs = ctx->vs;
 320    if (!vs->bo) {
 321       struct lima_screen *screen = lima_screen(ctx->base.screen);
 322       vs->bo = lima_bo_create(screen, vs->shader_size, 0);
 323       if (!vs->bo) {
 324          fprintf(stderr, "lima: create vs shader bo fail\n");
 325          return false;
 326       }
 327
 328       memcpy(lima_bo_map(vs->bo), vs->shader, vs->shader_size);
 329       ralloc_free(vs->shader);
 330       vs->shader = NULL;
 331    }
 332
 333    return true;
 334 }
 335
 336 bool
 337 lima_update_fs_state(struct lima_context *ctx)
 338 {
 339    struct lima_fs_shader_state *fs = ctx->fs;
 340    if (!fs->bo) {
 341       struct lima_screen *screen = lima_screen(ctx->base.screen);
 342       fs->bo = lima_bo_create(screen, fs->shader_size, 0);
 343       if (!fs->bo) {
 344          fprintf(stderr, "lima: create fs shader bo fail\n");
 345          return false;
 346       }
 347
 348       memcpy(lima_bo_map(fs->bo), fs->shader, fs->shader_size);
 349       ralloc_free(fs->shader);
 350       fs->shader = NULL;
 351    }
 352
 353    struct lima_job *job = lima_job_get(ctx);
 354    job->pp_max_stack_size = MAX2(job->pp_max_stack_size, ctx->fs->stack_size);
 355
 356    return true;
 357 }
 358
 359 static void *
 360 lima_create_vs_state(struct pipe_context *pctx,
 361                      const struct pipe_shader_state *cso)
 362 {
 363    struct lima_context *ctx = lima_context(pctx);
 364    struct lima_vs_shader_state *so = rzalloc(NULL, struct lima_vs_shader_state);
 365
 366    if (!so)
 367       return NULL;
 368
 369    nir_shader *nir;
 370    if (cso->type == PIPE_SHADER_IR_NIR)
 371       nir = cso->ir.nir;
 372    else {
 373       assert(cso->type == PIPE_SHADER_IR_TGSI);
 374
 375       nir = tgsi_to_nir(cso->tokens, pctx->screen);
 376    }
 377
 378    lima_program_optimize_vs_nir(nir);
 379
 380    if (lima_debug & LIMA_DEBUG_GP)
 381       nir_print_shader(nir, stdout);
 382
 383    if (!gpir_compile_nir(so, nir, &ctx->debug)) {
 384       ralloc_free(so);
 385       return NULL;
 386    }
 387
 388    ralloc_free(nir);
 389
 390    return so;
 391 }
 392
 393 static void
 394 lima_bind_vs_state(struct pipe_context *pctx, void *hwcso)
 395 {
 396    struct lima_context *ctx = lima_context(pctx);
 397
 398    ctx->vs = hwcso;
 399    ctx->dirty |= LIMA_CONTEXT_DIRTY_SHADER_VERT;
 400 }
 401
 402 static void
 403 lima_delete_vs_state(struct pipe_context *pctx, void *hwcso)
 404 {
 405    struct lima_vs_shader_state *so = hwcso;
 406
 407    if (so->bo)
 408       lima_bo_unreference(so->bo);
 409
 410    ralloc_free(so);
 411 }
 412
 413 void
 414 lima_program_init(struct lima_context *ctx)
 415 {
 416    ctx->base.create_fs_state = lima_create_fs_state;
 417    ctx->base.bind_fs_state = lima_bind_fs_state;
 418    ctx->base.delete_fs_state = lima_delete_fs_state;
 419
 420    ctx->base.create_vs_state = lima_create_vs_state;
 421    ctx->base.bind_vs_state = lima_bind_vs_state;
 422    ctx->base.delete_vs_state = lima_delete_vs_state;
 423 }