src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c

   1 /*
   2  * Copyright © 2019 Google, 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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   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 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 NONINFRINGEMENT.  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 FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  */
  23
  24 #include "ir3_nir.h"
  25 #include "ir3_compiler.h"
  26 #include "compiler/nir/nir.h"
  27 #include "compiler/nir/nir_builder.h"
  28 #include "util/u_math.h"
  29
  30 static inline struct ir3_ubo_range
  31 get_ubo_load_range(nir_intrinsic_instr *instr)
  32 {
  33         struct ir3_ubo_range r;
  34
  35         int offset = nir_src_as_uint(instr->src[1]);
  36         if (instr->intrinsic == nir_intrinsic_load_ubo_ir3)
  37                 offset *= 16;
  38         const int bytes = nir_intrinsic_dest_components(instr) * 4;
  39
  40         r.start = ROUND_DOWN_TO(offset, 16 * 4);
  41         r.end = ALIGN(offset + bytes, 16 * 4);
  42
  43         return r;
  44 }
  45
  46 static struct ir3_ubo_range *
  47 get_existing_range(nir_intrinsic_instr *instr,
  48                                    struct ir3_ubo_analysis_state *state,
  49                                    bool create_new)
  50 {
  51         unsigned block, base = 0;
  52         bool bindless;
  53         if (nir_src_is_const(instr->src[0])) {
  54                 block = nir_src_as_uint(instr->src[0]);
  55                 bindless = false;
  56         } else {
  57                 nir_intrinsic_instr *rsrc = ir3_bindless_resource(instr->src[0]);
  58                 if (rsrc && nir_src_is_const(rsrc->src[0])) {
  59                         block = nir_src_as_uint(rsrc->src[0]);
  60                         base = nir_intrinsic_desc_set(rsrc);
  61                         bindless = true;
  62                 } else {
  63                         return NULL;
  64                 }
  65         }
  66         for (int i = 0; i < IR3_MAX_UBO_PUSH_RANGES; i++) {
  67                 struct ir3_ubo_range *range = &state->range[i];
  68                 if (range->end < range->start) {
  69                         /* We don't have a matching range, but there are more available.
  70                          */
  71                         if (create_new) {
  72                                 range->block = block;
  73                                 range->bindless_base = base;
  74                                 range->bindless = bindless;
  75                                 return range;
  76                         } else {
  77                                 return NULL;
  78                         }
  79                 } else if (range->block == block && range->bindless_base == base &&
  80                                    range->bindless == bindless) {
  81                         return range;
  82                 }
  83         }
  84
  85         return NULL;
  86 }
  87
  88 static void
  89 gather_ubo_ranges(nir_shader *nir, nir_intrinsic_instr *instr,
  90                                   struct ir3_ubo_analysis_state *state)
  91 {
  92         struct ir3_ubo_range *old_r = get_existing_range(instr, state, true);
  93         if (!old_r)
  94                 return;
  95
  96         if (!nir_src_is_const(instr->src[1])) {
  97                 if (!old_r->bindless && old_r->block == 0) {
  98                         /* If this is an indirect on UBO 0, we'll still lower it back to
  99                          * load_uniform.  Set the range to cover all of UBO 0.
 100                          */
 101                         old_r->start = 0;
 102                         old_r->end = ALIGN(nir->num_uniforms * 16, 16 * 4);
 103                 }
 104
 105                 return;
 106         }
 107
 108         const struct ir3_ubo_range r = get_ubo_load_range(instr);
 109
 110         /* if UBO lowering is disabled, we still want to lower block 0
 111          * (which is normal uniforms):
 112          */
 113         if ((old_r->bindless || old_r->block != 0) && (ir3_shader_debug & IR3_DBG_NOUBOOPT))
 114                 return;
 115
 116         if (r.start < old_r->start)
 117                 old_r->start = r.start;
 118         if (old_r->end < r.end)
 119                 old_r->end = r.end;
 120 }
 121
 122 /* For indirect offset, it is common to see a pattern of multiple
 123  * loads with the same base, but different constant offset, ie:
 124  *
 125  *    vec1 32 ssa_33 = iadd ssa_base, const_offset
 126  *    vec4 32 ssa_34 = intrinsic load_uniform (ssa_33) (base=N, 0, 0)
 127  *
 128  * Detect this, and peel out the const_offset part, to end up with:
 129  *
 130  *    vec4 32 ssa_34 = intrinsic load_uniform (ssa_base) (base=N+const_offset, 0, 0)
 131  *
 132  * Or similarly:
 133  *
 134  *    vec1 32 ssa_33 = imad24_ir3 a, b, const_offset
 135  *    vec4 32 ssa_34 = intrinsic load_uniform (ssa_33) (base=N, 0, 0)
 136  *
 137  * Can be converted to:
 138  *
 139  *    vec1 32 ssa_base = imul24 a, b
 140  *    vec4 32 ssa_34 = intrinsic load_uniform (ssa_base) (base=N+const_offset, 0, 0)
 141  *
 142  * This gives the other opt passes something much easier to work
 143  * with (ie. not requiring value range tracking)
 144  */
 145 static void
 146 handle_partial_const(nir_builder *b, nir_ssa_def **srcp, int *offp)
 147 {
 148         if ((*srcp)->parent_instr->type != nir_instr_type_alu)
 149                 return;
 150
 151         nir_alu_instr *alu = nir_instr_as_alu((*srcp)->parent_instr);
 152
 153         if (alu->op == nir_op_imad24_ir3) {
 154                 /* This case is slightly more complicated as we need to
 155                  * replace the imad24_ir3 with an imul24:
 156                  */
 157                 if (!nir_src_is_const(alu->src[2].src))
 158                         return;
 159
 160                 *offp += nir_src_as_uint(alu->src[2].src);
 161                 *srcp = nir_imul24(b, nir_ssa_for_alu_src(b, alu, 0),
 162                                 nir_ssa_for_alu_src(b, alu, 1));
 163
 164                 return;
 165         }
 166
 167         if (alu->op != nir_op_iadd)
 168                 return;
 169
 170         if (!(alu->src[0].src.is_ssa && alu->src[1].src.is_ssa))
 171                 return;
 172
 173         if (nir_src_is_const(alu->src[0].src)) {
 174                 *offp += nir_src_as_uint(alu->src[0].src);
 175                 *srcp = alu->src[1].src.ssa;
 176         } else if (nir_src_is_const(alu->src[1].src)) {
 177                 *srcp = alu->src[0].src.ssa;
 178                 *offp += nir_src_as_uint(alu->src[1].src);
 179         }
 180 }
 181
 182 static void
 183 lower_ubo_block_decrement(nir_intrinsic_instr *instr, nir_builder *b)
 184 {
 185         /* Skip shifting things for turnip's bindless resources. */
 186         if (ir3_bindless_resource(instr->src[0]))
 187                 return;
 188
 189         /* Shift all GL nir_intrinsic_load_ubo UBO indices down by 1, because we
 190          * have lowered block 0 off of load_ubo to constbuf and ir3_const only
 191          * uploads pointers for block 1-N.
 192          */
 193         nir_ssa_def *old_idx = nir_ssa_for_src(b, instr->src[0], 1);
 194         nir_ssa_def *new_idx = nir_iadd_imm(b, old_idx, -1);
 195         nir_instr_rewrite_src(&instr->instr, &instr->src[0],
 196                         nir_src_for_ssa(new_idx));
 197 }
 198
 199 static void
 200 lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
 201                                                   struct ir3_ubo_analysis_state *state)
 202 {
 203         b->cursor = nir_before_instr(&instr->instr);
 204
 205         /* We don't lower dynamic block index UBO loads to load_uniform, but we
 206          * could probably with some effort determine a block stride in number of
 207          * registers.
 208          */
 209         struct ir3_ubo_range *range = get_existing_range(instr, state, false);
 210         if (!range) {
 211                 lower_ubo_block_decrement(instr, b);
 212                 return;
 213         }
 214
 215         if (range->bindless || range->block > 0) {
 216                 /* We don't lower dynamic array indexing either, but we definitely should.
 217                  * We don't have a good way of determining the range of the dynamic
 218                  * access, so for now just fall back to pulling.
 219                  */
 220                 if (!nir_src_is_const(instr->src[1])) {
 221                         lower_ubo_block_decrement(instr, b);
 222                         return;
 223                 }
 224
 225                 /* After gathering the UBO access ranges, we limit the total
 226                  * upload. Reject if we're now outside the range.
 227                  */
 228                 const struct ir3_ubo_range r = get_ubo_load_range(instr);
 229                 if (!(range->start <= r.start && r.end <= range->end)) {
 230                         lower_ubo_block_decrement(instr, b);
 231                         return;
 232                 }
 233         }
 234
 235         nir_ssa_def *ubo_offset = nir_ssa_for_src(b, instr->src[1], 1);
 236         int const_offset = 0;
 237
 238         handle_partial_const(b, &ubo_offset, &const_offset);
 239
 240         /* UBO offset is in bytes, but uniform offset is in units of
 241          * dwords, so we need to divide by 4 (right-shift by 2). For ldc the
 242          * offset is in units of 16 bytes, so we need to multiply by 4. And
 243          * also the same for the constant part of the offset:
 244          */
 245
 246         const int shift = instr->intrinsic == nir_intrinsic_load_ubo_ir3 ? 2 : -2;
 247         nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, ubo_offset, shift);
 248         nir_ssa_def *uniform_offset = NULL;
 249         if (new_offset) {
 250                 uniform_offset = new_offset;
 251         } else {
 252                 uniform_offset = shift > 0 ?
 253                         nir_ishl(b, ubo_offset, nir_imm_int(b,  shift)) :
 254                         nir_ushr(b, ubo_offset, nir_imm_int(b, -shift));
 255         }
 256
 257         if (instr->intrinsic == nir_intrinsic_load_ubo_ir3) {
 258                 const_offset <<= 2;
 259                 const_offset += nir_intrinsic_base(instr);
 260         } else {
 261                 debug_assert(!(const_offset & 0x3));
 262                 const_offset >>= 2;
 263         }
 264
 265         const int range_offset = ((int)range->offset - (int)range->start) / 4;
 266         const_offset += range_offset;
 267
 268         /* The range_offset could be negative, if if only part of the UBO
 269          * block is accessed, range->start can be greater than range->offset.
 270          * But we can't underflow const_offset.  If necessary we need to
 271          * insert nir instructions to compensate (which can hopefully be
 272          * optimized away)
 273          */
 274         if (const_offset < 0) {
 275                 uniform_offset = nir_iadd_imm(b, uniform_offset, const_offset);
 276                 const_offset = 0;
 277         }
 278
 279         nir_intrinsic_instr *uniform =
 280                 nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_uniform);
 281         uniform->num_components = instr->num_components;
 282         uniform->src[0] = nir_src_for_ssa(uniform_offset);
 283         nir_intrinsic_set_base(uniform, const_offset);
 284         nir_ssa_dest_init(&uniform->instr, &uniform->dest,
 285                                           uniform->num_components, instr->dest.ssa.bit_size,
 286                                           instr->dest.ssa.name);
 287         nir_builder_instr_insert(b, &uniform->instr);
 288         nir_ssa_def_rewrite_uses(&instr->dest.ssa,
 289                                                          nir_src_for_ssa(&uniform->dest.ssa));
 290
 291         nir_instr_remove(&instr->instr);
 292
 293         state->lower_count++;
 294 }
 295
 296 static bool
 297 instr_is_load_ubo(nir_instr *instr)
 298 {
 299         if (instr->type != nir_instr_type_intrinsic)
 300                 return false;
 301
 302         nir_intrinsic_op op = nir_instr_as_intrinsic(instr)->intrinsic;
 303         return op == nir_intrinsic_load_ubo || op == nir_intrinsic_load_ubo_ir3;
 304 }
 305
 306 bool
 307 ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader *shader)
 308 {
 309         struct ir3_ubo_analysis_state *state = &shader->ubo_state;
 310
 311         memset(state, 0, sizeof(*state));
 312         for (int i = 0; i < IR3_MAX_UBO_PUSH_RANGES; i++) {
 313                 state->range[i].start = UINT32_MAX;
 314         }
 315
 316         nir_foreach_function (function, nir) {
 317                 if (function->impl) {
 318                         nir_foreach_block (block, function->impl) {
 319                                 nir_foreach_instr (instr, block) {
 320                                         if (instr_is_load_ubo(instr))
 321                                                 gather_ubo_ranges(nir, nir_instr_as_intrinsic(instr), state);
 322                                 }
 323                         }
 324                 }
 325         }
 326
 327         /* For now, everything we upload is accessed statically and thus will be
 328          * used by the shader. Once we can upload dynamically indexed data, we may
 329          * upload sparsely accessed arrays, at which point we probably want to
 330          * give priority to smaller UBOs, on the assumption that big UBOs will be
 331          * accessed dynamically.  Alternatively, we can track statically and
 332          * dynamically accessed ranges separately and upload static rangtes
 333          * first.
 334          */
 335         const uint32_t max_upload = 16 * 1024;
 336         uint32_t offset = shader->const_state.num_reserved_user_consts * 16;
 337         state->num_enabled = ARRAY_SIZE(state->range);
 338         for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
 339                 if (state->range[i].start >= state->range[i].end) {
 340                         state->num_enabled = i;
 341                         break;
 342                 }
 343
 344                 uint32_t range_size = state->range[i].end - state->range[i].start;
 345
 346                 debug_assert(offset <= max_upload);
 347                 state->range[i].offset = offset;
 348                 if (offset + range_size > max_upload) {
 349                         range_size = max_upload - offset;
 350                         state->range[i].end = state->range[i].start + range_size;
 351                 }
 352                 offset += range_size;
 353
 354         }
 355         state->size = offset;
 356
 357         nir_foreach_function (function, nir) {
 358                 if (function->impl) {
 359                         nir_builder builder;
 360                         nir_builder_init(&builder, function->impl);
 361                         nir_foreach_block (block, function->impl) {
 362                                 nir_foreach_instr_safe (instr, block) {
 363                                         if (instr_is_load_ubo(instr))
 364                                                 lower_ubo_load_to_uniform(nir_instr_as_intrinsic(instr), &builder, state);
 365                                 }
 366                         }
 367
 368                         nir_metadata_preserve(function->impl, nir_metadata_block_index |
 369                                                                   nir_metadata_dominance);
 370                 }
 371         }
 372
 373         /* If we previously had UBO 0, it's been lowered off of load_ubo and all
 374          * the others were shifted down.
 375          */
 376         if (nir->info.num_ubos >= 1 && nir->info.first_ubo_is_default_ubo)
 377                 nir->info.num_ubos--;
 378
 379         return state->lower_count > 0;
 380 }