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